[
  {
    "id": "official-deployment",
    "contentType": "foundation",
    "category": "comfyui",
    "title": {
      "zh": "ComfyUI 从零到第一条 H3 带声视频",
      "en": "ComfyUI zero to your first H3 video with audio"
    },
    "outcome": {
      "zh": "从空白 ComfyUI 环境出发，正确放置 5 个模型文件并生成一条画面、对白或环境声都能播放的 T2VA MP4。",
      "en": "Start with a clean ComfyUI setup, place all five model files correctly, and render a playable T2VA MP4 with picture and synchronized audio."
    },
    "audience": {
      "zh": "第一次在 NVIDIA 电脑上本地运行 MiniMax H3、希望照着步骤完成而不是自己研究节点图的新手。",
      "en": "First-time NVIDIA users who want a guided local MiniMax H3 run instead of reverse-engineering a node graph."
    },
    "hardware": {
      "zh": "Windows 或 Linux + NVIDIA GPU；先准备充足磁盘空间。显存不足时先用模板预览尺寸，不要一开始就跑 1344×768。",
      "en": "Windows or Linux with an NVIDIA GPU and ample disk space. Start at the template preview size when VRAM is limited instead of jumping to 1344×768."
    },
    "prerequisites": {
      "zh": [
        "ComfyUI 0.30 或更新版本，能正常启动",
        "NVIDIA 驱动与 ComfyUI 安装方式匹配；桌面版按官方安装器处理依赖",
        "可访问模型文件，并为下载和输出预留空间"
      ],
      "en": [
        "ComfyUI 0.30 or newer, starting without errors",
        "NVIDIA drivers compatible with your ComfyUI installation; use the official installer for Desktop dependencies",
        "Access to model files and enough disk space for downloads and outputs"
      ]
    },
    "steps": {
      "zh": [
        "安装或更新 ComfyUI：Windows 新手先打开材料区的官方 Desktop 安装指南，按安装器选择 NVIDIA 并完成依赖安装，再用桌面应用启动。Linux/手动安装请先按官方手动指南克隆仓库、创建并激活独立 Python 环境和安装匹配 CUDA 的 PyTorch；只有这条路线才执行下方 pip 与 main.py 命令，工作目录为包含 requirements.txt 的 ComfyUI 仓库。确认页面能打开且没有 CUDA 初始化错误后再继续。",
        "打开 Template Library → Video → MiniMax H3 T2V。若弹出模型下载窗口，直接按提示下载；若手动下载，必须保持原文件名。",
        "把 minimax_h3_fl2va_pruned_int8_convrot.safetensors 放进 ComfyUI/models/diffusion_models/。T2V 与 I2V 都用 FL2VA；只有 R2V 才换成 Ref2VA 权重。",
        "把 qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors 放进 ComfyUI/models/text_encoders/；把 video VAE 和 audio VAE 一起放进 ComfyUI/models/vae/。缺少音频 VAE 时不能把静音结果当作成功。",
        "把 minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors 放进 ComfyUI/models/loras/。它用于模板的可选 turbo_mode；第一次基线先关闭 Turbo，使用默认 20 步。",
        "重启 ComfyUI 或刷新模型列表。逐个检查 Loader 下拉框能找到 diffusion model、text encoder、两个 VAE 与 LoRA；出现 null、undefined 或红色节点时先修路径，不要直接 Queue。",
        "先用模板预览尺寸、16:9、multiple=32 和短时长生成。Prompt 先写完整场景，再写分镜/运镜，最后写对白、音效和音乐；先不要加自定义节点或多种加速器。",
        "点击 Queue，等待 VAE 解码与视频封装完成。打开输出 MP4，分别检查画面、时长、声音和口型/事件同步，并保存这份工作流作为未修改的 baseline。",
        "基线通过后再把分辨率提高到 H3 原生 1344×768，或在 Resolution Selector 选择 0.98 MP；不要选 1.0 MP，它会得到超过面积上限的 1376×768。",
        "需要首尾帧时，在同一 FL2VA 工作流连接 first_frame/last_frame；需要人物、风格、动作或声音参考时，新建 R2V 模板并换成 Ref2VA 权重，不要在原图里混用两套模型。"
      ],
      "en": [
        "Install or update ComfyUI. Windows beginners should follow the linked official Desktop installer and launch the app after NVIDIA dependencies finish. For Linux/manual installs, first follow the manual guide to clone ComfyUI, activate its isolated Python environment and install the matching CUDA PyTorch build. Only that route uses the pip and main.py commands below, from the ComfyUI folder containing requirements.txt. Confirm the UI opens without CUDA errors before continuing.",
        "Open Template Library → Video → MiniMax H3 T2V. Use the model-download popup when available; preserve the exact filenames for manual downloads.",
        "Place minimax_h3_fl2va_pruned_int8_convrot.safetensors in ComfyUI/models/diffusion_models/. T2V and I2V use FL2VA; only R2V switches to Ref2VA weights.",
        "Place qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors in ComfyUI/models/text_encoders/. Put both the video and audio VAEs in ComfyUI/models/vae/. A silent render is not a successful audio-video setup.",
        "Place minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors in ComfyUI/models/loras/. It powers the optional turbo_mode; keep Turbo off and use the default 20 steps for the first baseline.",
        "Restart ComfyUI or refresh the model list. Confirm every Loader finds the diffusion model, text encoder, both VAEs, and LoRA. Fix null, undefined, or red nodes before queuing.",
        "Render a short clip at the template preview size, 16:9, and multiple=32. Describe the scene first, then shots and camera, then dialogue, effects, or music. Add no custom nodes or stacked accelerators yet.",
        "Queue the graph and wait through VAE decode and MP4 assembly. Open the file and verify picture, duration, audio, and event or lip synchronization. Save this untouched graph as the baseline.",
        "After the baseline passes, move to the native 1344×768 canvas or 0.98 MP in Resolution Selector. Do not choose 1.0 MP because it produces 1376×768, above the model area cap.",
        "For first/last frames, connect first_frame or last_frame in the same FL2VA graph. For identity, style, motion, camera, or voice references, open the R2V template and switch to Ref2VA weights instead of mixing model families."
      ]
    },
    "commands": [
      "python --version",
      "python -m pip install -r requirements.txt",
      "python main.py",
      "ComfyUI/models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors",
      "ComfyUI/models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
      "ComfyUI/models/vae/minimax_h3_video_vae_fp16.safetensors",
      "ComfyUI/models/vae/minimax_h3_audio_vae_fp32.safetensors",
      "ComfyUI/models/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors"
    ],
    "checks": {
      "zh": [
        "5 个模型文件都能在对应节点中选中，没有红色缺失节点",
        "20 步基线任务能完成，输出 MP4 可以站外播放器正常打开",
        "视频含声音，且对白、环境声或音效与画面事件基本同步",
        "已保存一份未叠加第三方节点的 baseline 工作流 JSON"
      ],
      "en": [
        "All five model files appear in their matching nodes with no red missing nodes",
        "The 20-step baseline finishes and its MP4 opens in an external player",
        "The video contains audio and dialogue, ambience, or effects broadly align with picture events",
        "An unmodified baseline workflow JSON is saved before adding third-party nodes"
      ]
    },
    "caveats": {
      "zh": [
        "模型文件很大；磁盘空间、下载中断和文件放错目录是最常见的首次失败原因。",
        "T2V/I2V 使用 FL2VA，R2V 使用 Ref2VA；权重混用会导致节点找不到模型或结果异常。",
        "H3 分辨率需为 32 的倍数；高分辨率和长时长会显著增加显存与等待时间。",
        "官方模板和文件名会更新；执行前用参考资料复核当前版本，不照抄过期截图。"
      ],
      "en": [
        "Model files are large; storage, interrupted downloads, and wrong folders are the most common first-run failures.",
        "T2V and I2V use FL2VA while R2V uses Ref2VA; mixing weight families causes missing models or bad output.",
        "H3 resolutions must be multiples of 32, and higher resolution or duration sharply increases VRAM and wait time.",
        "Templates and filenames evolve; verify the current reference before following an old screenshot."
      ]
    },
    "posterUrl": "/tutorial-posters/foundation-official.svg",
    "tags": [
      "ComfyUI",
      "T2VA",
      "Beginner"
    ],
    "relatedResourceIds": [
      "official-minimax-h3",
      "minimax-h3-easy",
      "huangserva-director-workflows",
      "comfyui-wiki"
    ],
    "source": {
      "platform": "docs",
      "url": "https://docs.comfy.org/tutorials/video/minimax/minimax-h3",
      "author": "ComfyUI",
      "authorProfileUrl": "https://github.com/Comfy-Org",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": true,
    "difficulty": "beginner",
    "hardwareProfiles": [
      "vram-16",
      "vram-24-plus",
      "cloud-gpu"
    ],
    "expectedResult": {
      "zh": "生成一条可在普通播放器打开、包含同步声音的短 MP4，并保存一份未叠加第三方节点的 baseline 工作流。",
      "en": "Render a short MP4 with synchronized audio that opens in a regular player, and save an untouched baseline workflow before adding third-party nodes."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "Loader 中找不到模型",
          "en": "A model is missing from a Loader"
        },
        "solution": {
          "zh": "核对五个文件的原始文件名与目录，刷新模型列表并彻底重启 ComfyUI。",
          "en": "Check the original filename and folder for all five files, refresh the model list, and fully restart ComfyUI."
        }
      },
      {
        "problem": {
          "zh": "输出静音或音画不同步",
          "en": "The output is silent or out of sync"
        },
        "solution": {
          "zh": "确认 audio VAE 已加载，并先在不含第三方节点的 20 步短片上复现。",
          "en": "Confirm the audio VAE is loaded and reproduce a short 20-step render without third-party nodes."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "删除本教程新增的模型文件前先确认没有其他工作流引用。",
        "保留 baseline JSON，出现兼容问题时退回官方模板。"
      ],
      "en": [
        "Before deleting downloaded model files, confirm no other workflow references them.",
        "Keep the baseline JSON and return to the official template when compatibility breaks."
      ]
    },
    "sourceRefs": [
      {
        "title": "MiniMax H3 official repository",
        "url": "https://github.com/MiniMax-AI/MiniMax-H3"
      }
    ],
    "learningTrack": "run",
    "depth": "deep",
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "official"
    },
    "applicableVersions": [
      "ComfyUI 0.30+ · native H3 T2V template"
    ],
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "使用官方原生模板，先建立一条带画面和声音的基线；不要求新手先学第三方节点。",
      "en": "The native template gives a picture-and-audio baseline before introducing third-party nodes."
    },
    "learningResources": [
      {
        "label": {
          "zh": "官方 T2V 工作流 JSON",
          "en": "Official T2V workflow JSON"
        },
        "url": "https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_t2v.json",
        "kind": "workflow"
      },
      {
        "label": {
          "zh": "Windows 官方 Desktop 安装步骤",
          "en": "Official Windows Desktop installation"
        },
        "url": "https://docs.comfy.org/installation/desktop/windows",
        "kind": "documentation"
      },
      {
        "label": {
          "zh": "Linux / 手动安装与 Python 环境",
          "en": "Linux / manual installation and Python environment"
        },
        "url": "https://docs.comfy.org/installation/manual_install",
        "kind": "documentation"
      },
      {
        "label": {
          "zh": "模型文件与目录",
          "en": "Model files and folders"
        },
        "url": "https://huggingface.co/Comfy-Org/MiniMax-H3",
        "kind": "models"
      },
      {
        "label": {
          "zh": "官方成品与 Prompt 示例",
          "en": "Official outputs and prompt examples"
        },
        "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md",
        "kind": "demo"
      }
    ],
    "cost": {
      "zh": "本地生成占用自己的硬件、内存和磁盘；模型与工具的使用条件请看原始说明。",
      "en": "Local generation uses your own hardware, RAM and storage. See the original model and tool terms."
    },
    "nextGuideIds": [
      "native-comfyui-modes",
      "context-ir-local-pipeline",
      "nvidia-comfyui"
    ],
    "relatedCases": [
      {
        "id": "official-t2va-starship",
        "title": {
          "zh": "舰桥上的跃迁余震",
          "en": "After the Fleet Jumps"
        },
        "relationship": "technique"
      }
    ],
    "learningOrder": 1,
    "guideType": "setup",
    "commandItems": [
      {
        "kind": "command",
        "value": "python --version",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "python -m pip install -r requirements.txt",
        "platform": "shell",
        "cwd": "ComfyUI"
      },
      {
        "kind": "command",
        "value": "python main.py",
        "platform": "shell",
        "cwd": "ComfyUI"
      },
      {
        "kind": "path",
        "value": "ComfyUI/models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors"
      },
      {
        "kind": "path",
        "value": "ComfyUI/models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors"
      },
      {
        "kind": "path",
        "value": "ComfyUI/models/vae/minimax_h3_video_vae_fp16.safetensors"
      },
      {
        "kind": "path",
        "value": "ComfyUI/models/vae/minimax_h3_audio_vae_fp32.safetensors"
      },
      {
        "kind": "path",
        "value": "ComfyUI/models/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors"
      }
    ],
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T09:41:09.754Z"
    }
  },
  {
    "id": "nvidia-comfyui",
    "contentType": "foundation",
    "category": "acceleration",
    "title": {
      "zh": "Turbo + 低显存：把 20 步降到 4–8 步",
      "en": "Turbo + lower VRAM: reduce 20 steps to 4–8"
    },
    "outcome": {
      "zh": "在已经跑通的 H3 ComfyUI 基线上安装社区 Turbo 节点，用 6–8 步获得可用成片，并在爆显存时有明确降级顺序。",
      "en": "Add the community Turbo nodes to a working H3 ComfyUI baseline, render usable clips at 6–8 steps, and follow a controlled fallback order when VRAM runs out."
    },
    "audience": {
      "zh": "已经能用官方 H3 工作流出片，但生成太慢或显存紧张的 NVIDIA 用户。",
      "en": "NVIDIA users who can already render with the official H3 graph but need faster iteration or lower VRAM use."
    },
    "hardware": {
      "zh": "NVIDIA GPU；先记录原生工作流的峰值显存与耗时。低显存模式是兜底，不是默认画质选项。",
      "en": "An NVIDIA GPU. Record native peak VRAM and time first. Low-VRAM mode is a fallback, not the default quality setting."
    },
    "prerequisites": {
      "zh": [
        "第一条教程的原生 T2V 或 I2V 工作流已经稳定输出",
        "固定 seed、Prompt、分辨率、时长和输入素材，便于 A/B 对比",
        "ComfyUI-Manager 可用，或能手动安装 custom_nodes",
        "从 Larryvrh 的 Hugging Face 仓库下载 minimax_h3_turbo_v4_step600_ema.safetensors"
      ],
      "en": [
        "A stable native T2V or I2V workflow from the first guide",
        "A fixed seed, prompt, resolution, duration, and input for A/B testing",
        "ComfyUI-Manager or the ability to install custom_nodes manually",
        "minimax_h3_turbo_v4_step600_ema.safetensors from Larryvrh's Hugging Face repository"
      ]
    },
    "steps": {
      "zh": [
        "先用原生 20 步工作流跑一次，记录总耗时、峰值显存、画面细节、快速运动和声音质量；这份结果是唯一基线。",
        "在 ComfyUI-Manager 搜索 MiniMax-H3 Turbo 并安装；或在 ComfyUI/custom_nodes/ 克隆仓库。完成后彻底重启 ComfyUI，确认控制台没有导入错误。",
        "把 minimax_h3_turbo_v4_step600_ema.safetensors 放入 ComfyUI/models/loras/。多数场景优先 v4-600；只有 4 步且高速大幅运动出现拖影时，才对比旧 v1-850。",
        "从官方 T2V/I2V 图开始，在 Load Diffusion Model 与 SamplerCustomAdvanced 之间插入 MiniMax-H3 Turbo LoRA 节点，并选择刚放入的权重。",
        "用 MiniMax-H3 Turbo Sampler 给 SamplerCustomAdvanced 提供 SAMPLER；BasicScheduler 设为 simple。其他 conditioning、双 VAE 解码和音频输出保持基线不变。",
        "第一轮设 8 步、strength=1.0；成功后再试 6 步。4 步只用于快速构图预览或明确的低预算场景，不要默认把 4 步结果当最终质量。",
        "用完全相同的 seed、Prompt、素材、分辨率和时长重新生成，比较耗时、脸手细节、快速运动拖影、音质与同步。一次只改 Turbo 这一个变量。",
        "如果爆显存，先降低预览分辨率或时长，再关闭非必要预览/节点，最后才打开 low_vram。low_vram 会降低峰值显存，但量化输出可能更软。",
        "如果画面拖影或模糊，可在明确问题片段把 strength 小幅提高到约 1.05–1.2；如果过锐和颗粒明显，降到约 0.8–0.95。每次只改一个参数。",
        "保留原生与 Turbo 两份工作流。Turbo 用于选镜与批量预览，重要镜头仍与 20 步原生基线对比后再决定成片版本。"
      ],
      "en": [
        "Run the native 20-step graph once and record total time, peak VRAM, fine detail, fast motion, and audio quality. This render is the only valid baseline.",
        "Install MiniMax-H3 Turbo through ComfyUI-Manager, or clone it under ComfyUI/custom_nodes/. Fully restart ComfyUI and confirm there is no import error.",
        "Place minimax_h3_turbo_v4_step600_ema.safetensors in ComfyUI/models/loras/. Prefer v4-600 for most work; compare the older v1-850 only when 4-step heavy motion shows trailing or smear.",
        "Start from the official T2V or I2V graph. Insert MiniMax-H3 Turbo LoRA between Load Diffusion Model and SamplerCustomAdvanced, then select the downloaded checkpoint.",
        "Feed SamplerCustomAdvanced from MiniMax-H3 Turbo Sampler and set BasicScheduler to simple. Keep conditioning, both VAE decoders, and audio output unchanged.",
        "Start at 8 steps and strength=1.0, then test 6 steps. Use 4 steps for composition previews or strict budgets rather than assuming it is final quality.",
        "Render again with the exact same seed, prompt, media, resolution, and duration. Compare time, face and hand detail, fast-motion trails, audio quality, and sync. Turbo must be the only changed variable.",
        "If VRAM runs out, reduce preview resolution or duration first, disable nonessential previews or nodes second, and enable low_vram last. Low-VRAM mode lowers peak use but can soften quantized output.",
        "For smear or blur on a specific clip, test strength around 1.05–1.2. For oversharp grain, test about 0.8–0.95. Change only one parameter per run.",
        "Keep separate native and Turbo workflow files. Use Turbo for shot selection and bulk previews, then compare important finals against the native 20-step baseline."
      ]
    },
    "commands": [
      "cd ComfyUI/custom_nodes",
      "git clone https://github.com/larryvrh/ComfyUI-MiniMax-H3-Turbo",
      "git -C ComfyUI-MiniMax-H3-Turbo pull"
    ],
    "checks": {
      "zh": [
        "MiniMax-H3 Turbo LoRA 与 Turbo Sampler 两个节点都能搜索并加载",
        "相同输入下，6–8 步 Turbo 成功输出带声 MP4",
        "已保存原生与 Turbo 的耗时、峰值显存和质量对比",
        "没有同时叠加 Spectrum、EasyCache 或其他未单独验证的加速器"
      ],
      "en": [
        "Both MiniMax-H3 Turbo LoRA and Turbo Sampler nodes load successfully",
        "The same input renders a video with audio at 6–8 Turbo steps",
        "Native and Turbo time, peak VRAM, and quality comparisons are saved",
        "Spectrum, EasyCache, and other untested accelerators are not stacked into this baseline"
      ]
    },
    "caveats": {
      "zh": [
        "4–8 步是有效范围，6–8 步通常明显优于 4 步；超过 8 步不再带来稳定收益，还可能过锐。",
        "strength 默认保持 1.0；只有具体片段出现拖影或过锐时才小幅调整。",
        "low_vram 会以清晰度换显存，不要在显存足够时默认开启。",
        "Turbo 的音质和高速运动仍在改进；不要把采样步数缩短直接宣传成等比例端到端加速。",
        "不要把 Turbo、Spectrum、EasyCache 等全部叠加；优化器的误差与音频损失可能累积。"
      ],
      "en": [
        "The useful range is 4–8 steps, with 6–8 usually much better than 4. Going above 8 adds no reliable benefit and may oversharpen.",
        "Keep strength at 1.0 by default and tune only for a specific smear or oversharp failure.",
        "low_vram trades sharpness for memory and should stay off when memory is sufficient.",
        "Audio and intense fast motion remain active areas of improvement. Fewer sampling steps are not the same as proportional end-to-end speedup.",
        "Do not stack Turbo, Spectrum, EasyCache, and other optimizers by default; their errors and audio losses can compound."
      ]
    },
    "posterUrl": "/tutorial-posters/foundation-comfyui.svg",
    "tags": [
      "Turbo",
      "Low VRAM",
      "6–8 Steps"
    ],
    "relatedResourceIds": [
      "turbo-lora",
      "modeltc-h3-turbo",
      "spectrum-h3",
      "minimax-h3-easy"
    ],
    "source": {
      "platform": "github",
      "url": "https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo",
      "author": "Larryvrh",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": true,
    "difficulty": "intermediate",
    "hardwareProfiles": [
      "vram-12",
      "vram-16",
      "vram-24-plus",
      "cloud-gpu"
    ],
    "expectedResult": {
      "zh": "在相同 seed、输入、分辨率和时长下得到 6–8 步带声视频，并留下原生 20 步与 Turbo 的耗时和质量对照。",
      "en": "Render a 6–8 step video with audio from the same seed, input, resolution, and duration, with a native 20-step comparison saved beside it."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "新节点没有出现在搜索中",
          "en": "The new nodes do not appear in search"
        },
        "solution": {
          "zh": "确认仓库位于 custom_nodes，查看启动日志中的导入错误并完全重启。",
          "en": "Confirm the repository is under custom_nodes, inspect startup import errors, and fully restart."
        }
      },
      {
        "problem": {
          "zh": "4 步画面拖影或声音变差",
          "en": "Four-step output smears or loses audio quality"
        },
        "solution": {
          "zh": "先回到 8 步、strength 1.0；不要叠加 Spectrum 或 EasyCache。",
          "en": "Return to eight steps at strength 1.0 and remove Spectrum or EasyCache from the comparison."
        }
      },
      {
        "problem": {
          "zh": "参考音频触发张量尺寸错误",
          "en": "Reference audio triggers a tensor-size mismatch"
        },
        "solution": {
          "zh": "先查看 Issue #27 与已合并的 PR #16，更新 Turbo 节点并完整重启。仍失败时回到不加 Turbo 的官方参考工作流；不要把 low_vram 开关当作已证实的修复。",
          "en": "Check Issue #27 and merged PR #16, update the Turbo nodes and fully restart. If it persists, return to the official reference workflow without Turbo; low_vram is not an established fix."
        }
      },
      {
        "problem": {
          "zh": "显存不足，任务中止",
          "en": "Out of VRAM: generation stops"
        },
        "solution": {
          "zh": "先降低分辨率与帧数，再按 Turbo README 开启 low_vram；量化基模合并 LoRA 后可能更软，应与原生基线比较。",
          "en": "Reduce resolution and frame count, then enable low_vram as described in the Turbo README. Merging LoRA into a quantized base can soften output; compare against the native baseline."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "删除 ComfyUI/custom_nodes/ComfyUI-MiniMax-H3-Turbo 并重启。",
        "移除工作流中的 Turbo LoRA 与 Turbo Sampler 节点即可回到原生采样。"
      ],
      "en": [
        "Delete ComfyUI/custom_nodes/ComfyUI-MiniMax-H3-Turbo and restart.",
        "Remove the Turbo LoRA and Turbo Sampler nodes to restore native sampling."
      ]
    },
    "sourceRefs": [
      {
        "title": "Turbo LoRA weights",
        "url": "https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora"
      },
      {
        "title": "Reference-audio issue and repair discussion #27",
        "url": "https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo/issues/27"
      }
    ],
    "learningTrack": "run",
    "depth": "deep",
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "author-docs",
      "communityReviewedAt": "2026-09-06"
    },
    "applicableVersions": [
      "Larryvrh H3 Turbo · v4 step600 EMA",
      "Native template Turbo and plugin Turbo use different setup paths"
    ],
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "把低显存和加速分开判断：先有基线，再单独测试 Turbo；视频提供节点演示，仓库说明插件的参数。",
      "en": "Separate memory limits from speed: establish a baseline, then test Turbo alone. The video demonstrates nodes and the repository documents plugin settings."
    },
    "learningResources": [
      {
        "label": {
          "zh": "Turbo 插件与工作流说明",
          "en": "Turbo plugin and workflow instructions"
        },
        "url": "https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo/tree/main/example_workflows",
        "kind": "workflow"
      },
      {
        "label": {
          "zh": "Turbo LoRA 权重",
          "en": "Turbo LoRA weights"
        },
        "url": "https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora",
        "kind": "models"
      },
      {
        "label": {
          "zh": "观看作者分章演示",
          "en": "Watch the chaptered video tutorial"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g",
        "kind": "demo"
      },
      {
        "label": {
          "zh": "官方原生 Turbo 对照",
          "en": "Official native Turbo reference"
        },
        "url": "https://docs.comfy.org/tutorials/video/minimax/minimax-h3",
        "kind": "documentation"
      }
    ],
    "cost": {
      "zh": "本地生成占用自己的硬件、内存和磁盘；模型与工具的使用条件请看原始说明。",
      "en": "Local generation uses your own hardware, RAM and storage. See the original model and tool terms."
    },
    "nextGuideIds": [
      "official-deployment",
      "prompt-agent-skill"
    ],
    "chapters": [
      {
        "title": {
          "zh": "Turbo 加速",
          "en": "Turbo speedup"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=480",
        "seconds": 480
      },
      {
        "title": {
          "zh": "低显存路线",
          "en": "Lower VRAM"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=699",
        "seconds": 699
      },
      {
        "title": {
          "zh": "简化工作流",
          "en": "Simpler workflow"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=942",
        "seconds": 942
      },
      {
        "title": {
          "zh": "音频生成",
          "en": "Audio generation"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=1148",
        "seconds": 1148
      },
      {
        "title": {
          "zh": "Prompt 写法",
          "en": "Prompting"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=1380",
        "seconds": 1380
      },
      {
        "title": {
          "zh": "GGUF 量化选择",
          "en": "GGUF quantization options"
        },
        "seconds": 1571,
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=1571"
      }
    ],
    "communityFeedback": [
      {
        "kind": "usage",
        "summary": {
          "zh": "原视频评论有人反馈预览节点帮助提前判断是否值得继续等待；这属于使用体验，不能当作固定加速倍数。",
          "en": "A viewer reports that previews help decide whether a render is worth waiting for. This is usability feedback, not a measured speed multiplier."
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&lc=UgyTbfwka0nzc96pWiZ4AaABAg"
      },
      {
        "kind": "praise",
        "summary": {
          "zh": "评论对工作流讲解和资源整理有正面反馈；其中“准备尝试”类留言不能算复现成功。",
          "en": "Comments praise the workflow explanations and collected resources. Comments saying they plan to try it are not successful reproductions."
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&lc=UgzY0sgAQO3qR4zQHmN4AaABAg"
      },
      {
        "kind": "issue",
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&lc=UgwQnHAYkAeXotwLTYB4AaABAg",
        "summary": {
          "zh": "有观众报告生成一直停在 0%，但未提供硬件、日志或版本，所见评论没有作者解决答复；不能据此确定原因或保证所有机器可运行。",
          "en": "A viewer reports generation stuck at 0%, without hardware, logs or versions. No author resolution was visible for that comment; it does not establish a cause or universal compatibility."
        }
      },
      {
        "kind": "fix",
        "url": "https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo/issues/27",
        "summary": {
          "zh": "Issue #27 记录了 ComfyUI 0.31.1、RTX 5090、pruned Ref2VA 配合参考音频时的张量尺寸错误。修复贡献者指向已合并的 PR #16；原问题仍开放，更新后仍应先做短片检查，不直接叠加参考与续接。",
          "en": "Issue #27 reports a tensor-size mismatch with ComfyUI 0.31.1, RTX 5090 and pruned Ref2VA plus reference audio. A contributor points to merged PR #16; the issue remains open. Update and check a short clip before combining references or continuation."
        }
      }
    ],
    "learningOrder": 4,
    "guideType": "project",
    "commandItems": [
      {
        "kind": "command",
        "value": "cd ComfyUI/custom_nodes",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "git clone https://github.com/larryvrh/ComfyUI-MiniMax-H3-Turbo",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "git -C ComfyUI-MiniMax-H3-Turbo pull",
        "platform": "shell"
      }
    ],
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T09:41:10.975Z"
    }
  },
  {
    "id": "mac-native",
    "contentType": "foundation",
    "category": "getting-started",
    "title": {
      "zh": "Mac 从零运行 H3：纯 C + Metal",
      "en": "Run H3 on Mac from zero with pure C + Metal"
    },
    "outcome": {
      "zh": "不用 Python、PyTorch 或 ComfyUI，在 Apple Silicon 上编译 h3.c、检查模型，并生成第一条含声音的 MP4。",
      "en": "Without Python, PyTorch, or ComfyUI, compile h3.c on Apple Silicon, validate the model, and generate a first MP4 with audio."
    },
    "audience": {
      "zh": "有 Apple Silicon Mac、能使用终端，但不想搭建 NVIDIA/ComfyUI 环境的用户。",
      "en": "Apple Silicon users who can use Terminal and do not want to build an NVIDIA or ComfyUI stack."
    },
    "hardware": {
      "zh": "Apple Silicon Mac。作者常驻 BF16 路径的 DiT 约占 37 GiB，另需文本编码、VAE 和系统余量；内存不足时按 README 选 SSD streaming，不把 DiT 的占用当整机最低内存。先从 512×512、22 帧开始。",
      "en": "Apple Silicon Mac. The author’s resident BF16 DiT uses about 37 GiB, plus headroom for text encoding, VAEs and macOS. Use the README’s SSD-streaming path if memory is tight; DiT storage is not the minimum total system RAM. Start at 512×512 and 22 frames."
    },
    "prerequisites": {
      "zh": [
        "macOS、Git 与 Xcode Command Line Tools",
        "FFmpeg 和 FFprobe 可从终端直接调用",
        "先安装 Hugging Face CLI：brew install hf；运行 hf --help 确认可用，再下载完整模型到 h3.c/MiniMax-H3",
        "关闭占用大量统一内存的应用，并为首次载入预留足够时间"
      ],
      "en": [
        "macOS, Git, and Xcode Command Line Tools",
        "FFmpeg and FFprobe callable from Terminal",
        "Install the Hugging Face CLI with brew install hf; check hf --help before downloading the full model into h3.c/MiniMax-H3",
        "Memory-heavy applications closed and enough time reserved for the first load"
      ]
    },
    "steps": {
      "zh": [
        "打开终端运行 xcode-select --install；若已安装会提示无需重复。安装 Homebrew 后执行 brew install ffmpeg，再分别检查 ffmpeg 和 ffprobe 版本。",
        "克隆 antirez/h3.c 并进入目录。不要把仓库和模型权重混在同一层散放：教程默认模型完整位于 ./MiniMax-H3。",
        "从 Hugging Face 下载 MiniMaxAI/MiniMax-H3 的完整 snapshot 到 ./MiniMax-H3。下载完成后不要改权重文件名，也不要只下载 README 中看起来像主模型的单个文件。",
        "执行 make -j8 编译；创建 outputs 目录。若编译失败，先确认当前终端使用的是 Apple clang，并重新检查 Command Line Tools，而不是随意删除编译参数。",
        "运行 ./h3 --info -d ./MiniMax-H3。它只核验模型布局并打印 Metal 设备，不会映射全部权重或开始生成；任何缺失文件都应在这一步解决。",
        "先运行 512×512、22 帧、20 步的平衡示例。首次启动包含模型载入与文件缓存，不能用第一次总耗时代表稳定速度。",
        "只有终端支持图像预览时才添加 --show；内存紧张时省略它。用 --profile 记录各阶段耗时，勿把其他芯片的成绩当成本机保证。",
        "确认 outputs/fox-fast.mp4 能播放且含声音。再以相同命令重复一次，第二次耗时才更适合与其他设置比较。",
        "需要交互创作时运行 ./h3 -d ./MiniMax-H3 --width 512 --height 512 --steps 6。使用 !status、!seed random、!seconds 2 与 !save output.mp4 管理会话。",
        "首尾帧使用 !first opening.png 与 !last ending.png；一般参考图使用 !ref-image person.png。Ref2VA 参考图不能和 !first/!last 同时使用，切换模式前先清理旧条件。"
      ],
      "en": [
        "Run xcode-select --install in Terminal; macOS will say when it is already present. Install FFmpeg with Homebrew, then verify both ffmpeg and ffprobe.",
        "Clone antirez/h3.c and enter the repository. Keep the code separate from the weight files; this guide expects the complete model at ./MiniMax-H3.",
        "Download the complete MiniMaxAI/MiniMax-H3 Hugging Face snapshot into ./MiniMax-H3. Preserve filenames and do not download only the apparent main checkpoint.",
        "Run make -j8, then create outputs. If compilation fails, confirm the active compiler is Apple clang and repair Command Line Tools before changing build flags.",
        "Run ./h3 --info -d ./MiniMax-H3. It validates model layout and prints the Metal device without mapping every weight or generating media. Resolve missing files here.",
        "Run the balanced 512×512, 22-frame, 20-step example first. The first invocation pays model-loading and filesystem-cache costs, so its total time is not a stable benchmark.",
        "Add --show only in a terminal that supports image previews; omit it under memory pressure. Use --profile for phase timings and do not treat results from another chip as a guarantee.",
        "Confirm outputs/fox-fast.mp4 plays with audio. Repeat the same command; the second run is a more useful performance comparison.",
        "For interactive work, run ./h3 -d ./MiniMax-H3 --width 512 --height 512 --steps 6. Use !status, !seed random, !seconds 2, and !save output.mp4 to control the session.",
        "Use !first opening.png and !last ending.png for boundary frames, or !ref-image person.png for a general reference. Ref2VA references cannot be mixed with !first or !last; clear the old mode first."
      ]
    },
    "commands": [
      "xcode-select --install",
      "brew install ffmpeg",
      "git clone https://github.com/antirez/h3.c && cd h3.c",
      "brew install hf",
      "hf --help",
      "hf download MiniMaxAI/MiniMax-H3 --local-dir MiniMax-H3",
      "make -j8 && mkdir -p outputs",
      "./h3 --info -d ./MiniMax-H3",
      "./h3 --profile -d ./MiniMax-H3 -p \"A red fox walks through fresh snow in a pine forest. Medium tracking shot, natural winter light, realistic fur, soft footsteps and wind.\" --width 512 --height 512 --frames 22 --steps 20 --layers 45 --reuse 2 -o outputs/fox-fast.mp4",
      "./h3 -d ./MiniMax-H3 --width 512 --height 512 --steps 6",
      "make test"
    ],
    "checks": {
      "zh": [
        "./h3 --info 能识别完整模型目录并打印 Apple Metal 设备",
        "outputs/fox-fast.mp4 可播放、含画面和声音，时长约 0.92 秒",
        "相同命令第二次运行成功，并记录了不含首次载入开销的耗时",
        "交互模式能接受 Prompt，并能用 !save 保存指定文件"
      ],
      "en": [
        "./h3 --info accepts the complete model directory and prints the Apple Metal device",
        "outputs/fox-fast.mp4 plays with picture and audio for roughly 0.92 seconds",
        "A second identical run succeeds and its time excludes most first-load overhead",
        "Interactive mode accepts a prompt and !save writes the requested file"
      ]
    },
    "caveats": {
      "zh": [
        "h3.c 正在快速迭代；权重布局、命令参数和性能结论必须以当前 README 为准。",
        "--show 是可选预览，不是生成所必需；它会增加解码时间和约 10 GiB 临时内存占用。",
        "4 步适合极快预览，但细节与运动质量低于 20 步；不要把不同步数结果当同质量比较。",
        "Apple Silicon 使用统一内存；系统剩余内存不足时，交换与热降频会让结果远慢于公开基准。",
        "低内存可按原作使用 --ssd-streaming，并省略 --show；这是内存与速度的取舍，不能与 --use-int8-row-fc2 组合。"
      ],
      "en": [
        "h3.c evolves quickly; verify weight layout, CLI flags, and performance claims against the current README.",
        "--show is an optional preview, not a generation requirement; it adds decode time and about 10 GiB of temporary memory.",
        "Four steps are useful for very fast previews but lose detail and motion quality versus 20; do not compare them as equal-quality runs.",
        "Apple Silicon uses unified memory. Low free memory, swapping, and thermal throttling can make a run far slower than public benchmarks.",
        "For low memory, follow the source’s --ssd-streaming path and omit --show. This trades speed for memory and cannot be combined with --use-int8-row-fc2."
      ]
    },
    "posterUrl": "/tutorial-posters/foundation-mac.svg",
    "tags": [
      "Apple Silicon",
      "Metal",
      "0→1"
    ],
    "relatedResourceIds": [
      "antirez-h3-metal",
      "awesome-h3-resources"
    ],
    "source": {
      "platform": "github",
      "url": "https://github.com/antirez/h3.c",
      "author": "antirez",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": true,
    "difficulty": "intermediate",
    "hardwareProfiles": [
      "apple-silicon"
    ],
    "expectedResult": {
      "zh": "h3 --info 识别模型与 Metal 设备，并生成 outputs/fox-fast.mp4；文件包含可解码的视频轨和声音轨。",
      "en": "h3 --info identifies the model and Metal device, and outputs/fox-fast.mp4 contains decodable video and audio tracks."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "编译器或 Metal 构建失败",
          "en": "Compiler or Metal build failure"
        },
        "solution": {
          "zh": "重新运行 xcode-select --install，确认使用 Apple clang，再执行 make clean 后重新编译。",
          "en": "Run xcode-select --install, confirm Apple clang is active, then run make clean and rebuild."
        }
      },
      {
        "problem": {
          "zh": "--info 报模型文件缺失",
          "en": "--info reports missing model files"
        },
        "solution": {
          "zh": "重新核对完整 Hugging Face snapshot，不要只下载主权重或修改文件名。",
          "en": "Recheck the complete Hugging Face snapshot; do not download only the main checkpoint or rename files."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "删除 h3.c 工作目录即可移除程序；模型目录可独立保留。",
        "用 brew uninstall ffmpeg 移除本教程安装的 FFmpeg，但先确认没有其他项目依赖。"
      ],
      "en": [
        "Remove the h3.c working directory to uninstall the program; the model directory can be kept separately.",
        "Use brew uninstall ffmpeg only after confirming no other project depends on it."
      ]
    },
    "sourceRefs": [
      {
        "title": "MiniMax H3 weights",
        "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3"
      }
    ],
    "learningTrack": "run",
    "depth": "deep",
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "author-docs"
    },
    "applicableVersions": [
      "h3.c README · 2026-09-06",
      "Apple Silicon · FFmpeg / FFprobe"
    ],
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "适合偏好终端的 Mac 用户：原作者提供纯 C + Metal 路线和最小输出命令，可先检查模型目录再生成。",
      "en": "For Mac users comfortable with Terminal: the author provides a C + Metal route and a small output command, with model inspection before generation."
    },
    "learningResources": [
      {
        "label": {
          "zh": "h3.c 安装、命令与示例",
          "en": "h3.c installation, commands and examples"
        },
        "url": "https://github.com/antirez/h3.c",
        "kind": "documentation"
      },
      {
        "label": {
          "zh": "原始权重与模型说明",
          "en": "Original weights and model documentation"
        },
        "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3",
        "kind": "models"
      }
    ],
    "cost": {
      "zh": "本地生成占用自己的硬件、内存和磁盘；模型与工具的使用条件请看原始说明。",
      "en": "Local generation uses your own hardware, RAM and storage. See the original model and tool terms."
    },
    "nextGuideIds": [
      "context-ir-local-pipeline",
      "native-comfyui-modes"
    ],
    "learningOrder": 2,
    "guideType": "setup",
    "commandItems": [
      {
        "kind": "command",
        "value": "xcode-select --install",
        "platform": "macos"
      },
      {
        "kind": "command",
        "value": "brew install ffmpeg",
        "platform": "macos"
      },
      {
        "kind": "command",
        "value": "git clone https://github.com/antirez/h3.c && cd h3.c",
        "platform": "macos"
      },
      {
        "kind": "command",
        "value": "brew install hf",
        "platform": "macos"
      },
      {
        "kind": "command",
        "value": "hf --help",
        "platform": "macos"
      },
      {
        "kind": "command",
        "value": "hf download MiniMaxAI/MiniMax-H3 --local-dir MiniMax-H3",
        "platform": "macos",
        "cwd": "h3.c"
      },
      {
        "kind": "command",
        "value": "make -j8 && mkdir -p outputs",
        "platform": "macos",
        "cwd": "h3.c"
      },
      {
        "kind": "command",
        "value": "./h3 --info -d ./MiniMax-H3",
        "platform": "macos",
        "cwd": "h3.c"
      },
      {
        "kind": "command",
        "value": "./h3 --profile -d ./MiniMax-H3 -p \"A red fox walks through fresh snow in a pine forest. Medium tracking shot, natural winter light, realistic fur, soft footsteps and wind.\" --width 512 --height 512 --frames 22 --steps 20 --layers 45 --reuse 2 -o outputs/fox-fast.mp4",
        "platform": "macos",
        "cwd": "h3.c"
      },
      {
        "kind": "command",
        "value": "./h3 -d ./MiniMax-H3 --width 512 --height 512 --steps 6",
        "platform": "macos",
        "cwd": "h3.c"
      },
      {
        "kind": "command",
        "value": "make test",
        "platform": "macos",
        "cwd": "h3.c"
      }
    ],
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T09:41:12.984Z"
    }
  },
  {
    "id": "prompt-agent-skill",
    "contentType": "foundation",
    "category": "long-video",
    "title": {
      "zh": "Motion Context：把两段 H3 接成长视频",
      "en": "Motion Context: chain H3 clips into a longer video"
    },
    "outcome": {
      "zh": "安装 Motion Context 节点，用上一段的音画 latent 续接下一段，完成一次画面与声音都连续的两段拼接。",
      "en": "Install Motion Context, carry the previous audio-video latent into the next generation, and complete a two-clip chain with continuous picture and sound."
    },
    "audience": {
      "zh": "已经能在 ComfyUI 稳定生成单段 H3，想突破单次时长而不接受明显画面接缝或声音重启的用户。",
      "en": "ComfyUI users with a stable single-clip H3 workflow who need longer sequences without obvious picture seams or restarted audio."
    },
    "hardware": {
      "zh": "能稳定运行 H3 的 NVIDIA 环境；两段必须保持完全相同的分辨率，建议先用短片和原生采样验证。",
      "en": "An NVIDIA setup that already runs H3 reliably. Both clips must use identical resolution; validate with short native-sampling clips first."
    },
    "prerequisites": {
      "zh": [
        "先有可播放的 H3 基线工作流",
        "当前主线使用 ComfyUI 0.34+；旧环境按 CHANGELOG 选择 v0.3.1",
        "为中间 latent 和视频输出预留磁盘空间"
      ],
      "en": [
        "A working H3 baseline with playable output",
        "Use ComfyUI 0.34+ for the current main branch; select v0.3.1 for older environments as described in CHANGELOG",
        "Disk space for intermediate latents and output videos"
      ]
    },
    "steps": {
      "zh": [
        "先检查 ComfyUI 版本并安装兼容的 Motion Context；重启后确认节点注册，首次生成时检查 layout checks 日志。",
        "复制基线或打开示例 JSON，把 Motion Context 接到 conditioning 路径；Save Latent 接采样器输出，Load Latent 接 context_latent。",
        "把 Load、Save 和 Chain 放进同一个画布分组。第一段设 Load=0、Save=1；Load=0 不读文件，可让第一段直接通过。",
        "使用 Chain 节点的 Run/Re-roll 生成第一段。检查输出后，满意再 Approve；它会同时推进索引并排队下一段。",
        "第二段沿用相同尺寸，用前一段结尾的构图和动作开始新 Prompt。不要一开始就换场景，以免难以判断续接问题。",
        "让解码后的 IMAGE 与 AUDIO 一起经过 Trim，并接入 trim_frames；输出会裁去用于衔接的重叠头部。",
        "播放拼接点，分别检查重复帧、运动方向、静帧和声音重启。通过后再扩展第三段，并保存工作流与片段索引。"
      ],
      "en": [
        "Check the ComfyUI version and install compatible Motion Context. After restarting, confirm node registration and inspect layout checks on first use.",
        "Duplicate the baseline or open the example JSON. Put Motion Context in the conditioning path, Save Latent after the sampler, and Load Latent into context_latent.",
        "Place Load, Save and Chain in one canvas group. For clip one use Load=0 and Save=1; Load=0 reads no file and passes the first clip through.",
        "Generate clip one with the Chain node Run/Re-roll control. Inspect it, then use Approve to advance both indices and queue the next clip.",
        "Keep the same dimensions for clip two and begin its prompt with the preceding composition and motion. Avoid an immediate location change while evaluating continuity.",
        "Pass decoded IMAGE and AUDIO through Trim and connect trim_frames. It removes the overlapping head used for continuity.",
        "Play across the join and inspect repeated frames, motion direction, freezes and restarted sound. Extend to clip three only after checking the join, and save the workflow and indices."
      ]
    },
    "commands": [
      "cd ComfyUI/custom_nodes",
      "git clone https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context"
    ],
    "checks": {
      "zh": [
        "Load、Save 与 Chain 在同一组；索引与实际片段一致。",
        "Trim 同时处理画面和声音，没有重复的衔接头部。",
        "拼接点经过视听检查；不能仅凭脚本执行完毕宣称连续。"
      ],
      "en": [
        "Load, Save and Chain share a group and indices match the actual clips.",
        "Trim handles both picture and sound without a repeated overlap head.",
        "The join has been watched and listened to; successful execution alone does not prove continuity."
      ]
    },
    "caveats": {
      "zh": [
        "不要用 queue run on change 推进链条；索引修改可能触发重复排队。",
        "软件布局检查失败时先看版本说明，不能把测试通过等同于生成效果通过。"
      ],
      "en": [
        "Do not advance a chain with queue run on change; index changes can trigger duplicate runs.",
        "If layout checks fail, consult compatibility notes. Passing code tests does not prove output quality."
      ]
    },
    "posterUrl": "/tutorial-posters/foundation-prompt.svg",
    "tags": [
      "Motion Context",
      "Long Video",
      "Audio"
    ],
    "relatedResourceIds": [
      "motion-context",
      "audio-t8",
      "official-minimax-h3"
    ],
    "source": {
      "platform": "github",
      "url": "https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context",
      "author": "NikoDemon80",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": true,
    "difficulty": "advanced",
    "hardwareProfiles": [
      "vram-16",
      "vram-24-plus",
      "cloud-gpu"
    ],
    "expectedResult": {
      "zh": "第二段延续第一段的运动方向与声音，并在 Trim 后移除重复上下文；两段分辨率一致、接缝可检查。",
      "en": "Clip B continues clip A's motion direction and sound, duplicate context is removed after Trim, and both clips share one resolution with an inspectable join."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "Load Latent 形成循环连接",
          "en": "Load Latent creates a circular graph"
        },
        "solution": {
          "zh": "使用插件自带的 Save/Load Latent 跨运行传递，并明确设置上一段与当前段的 clip_index。",
          "en": "Use the pack's Save/Load Latent nodes across runs and set the previous/current clip_index explicitly."
        }
      },
      {
        "problem": {
          "zh": "接缝重复或音频错位",
          "en": "The join repeats or audio drifts"
        },
        "solution": {
          "zh": "把 trim_frames 接到 Motion Context Trim，保持 match_tail，并从 22 帧视频、24 帧音频上下文重新测试。",
          "en": "Wire trim_frames into Motion Context Trim, keep match_tail enabled, and retry with 22 video and 24 audio context frames."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "删除 ComfyUI/custom_nodes/ComfyUI-H3-Motion-Context 并重启。",
        "移除 Motion Context、Save/Load Latent 与 Trim 节点后恢复普通 H3 图。"
      ],
      "en": [
        "Delete ComfyUI/custom_nodes/ComfyUI-H3-Motion-Context and restart.",
        "Remove Motion Context, Save/Load Latent, and Trim nodes to restore a standard H3 graph."
      ]
    },
    "sourceRefs": [
      {
        "title": "ComfyUI native H3 guide",
        "url": "https://docs.comfy.org/tutorials/video/minimax/minimax-h3"
      }
    ],
    "learningTrack": "create",
    "depth": "deep",
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "author-docs"
    },
    "applicableVersions": [
      "ComfyUI 0.34+ · current Motion Context",
      "ComfyUI 0.33.4 or older: use Motion Context v0.3.1"
    ],
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "先练习两段的画面与声音衔接，再扩展短片；保留 latent、正确裁去重叠帧，比只拼接 MP4 更接近连续生成。",
      "en": "Practice a two-clip picture-and-sound join before extending a short film. Carrying latents and trimming overlap supports continuity beyond simply concatenating MP4s."
    },
    "learningResources": [
      {
        "label": {
          "zh": "Motion Context 安装与连接说明",
          "en": "Motion Context installation and wiring"
        },
        "url": "https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context",
        "kind": "documentation"
      },
      {
        "label": {
          "zh": "原作者示例工作流",
          "en": "Author example workflows"
        },
        "url": "https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context/tree/main/example_workflows",
        "kind": "workflow"
      },
      {
        "label": {
          "zh": "版本变更与旧版兼容",
          "en": "Version changes and compatibility"
        },
        "url": "https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context/blob/main/CHANGELOG.md",
        "kind": "documentation"
      }
    ],
    "cost": {
      "zh": "本地生成占用自己的硬件、内存和磁盘；模型与工具的使用条件请看原始说明。",
      "en": "Local generation uses your own hardware, RAM and storage. See the original model and tool terms."
    },
    "nextGuideIds": [
      "video-pov-camera-angles",
      "nvidia-comfyui"
    ],
    "learningOrder": 4,
    "guideType": "project",
    "commandItems": [
      {
        "kind": "command",
        "value": "cd ComfyUI/custom_nodes",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "git clone https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context",
        "platform": "shell"
      }
    ],
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T09:41:14.408Z"
    }
  },
  {
    "id": "local-deployment-zero-to-output",
    "contentType": "community",
    "category": "getting-started",
    "title": {
      "zh": "从零装机到第一条带声视频",
      "en": "From a clean machine to the first video with audio"
    },
    "outcome": {
      "zh": "建立一条可复现的本地 H3 部署与 Prompt 起步流程。",
      "en": "Build a reproducible local H3 setup and prompt-writing baseline."
    },
    "audience": {
      "zh": "第一次部署 H3 的 NVIDIA 用户。",
      "en": "NVIDIA users deploying H3 for the first time."
    },
    "hardware": {
      "zh": "NVIDIA GPU；具体显存与软件版本按原帖文章及官方 README 复核。",
      "en": "An NVIDIA GPU; recheck VRAM and software versions against the source article and official README."
    },
    "prerequisites": {
      "zh": [
        "系统和显卡驱动可用",
        "准备模型下载空间",
        "能运行 ComfyUI"
      ],
      "en": [
        "A working OS and GPU driver",
        "Enough storage for model downloads",
        "A runnable ComfyUI installation"
      ]
    },
    "steps": {
      "zh": [
        "按原帖文章准备依赖与模型目录。",
        "先复现文中的带声视频基线。",
        "再按官方 Prompt 语法替换镜头、对白和环境声。"
      ],
      "en": [
        "Prepare dependencies and model directories from the source article.",
        "Reproduce its audio-video baseline first.",
        "Then replace shot, dialogue, and ambience using the official prompt syntax."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "执行前再次核验官方 README；不要把文章中的旧版本号当成永久要求。"
      ],
      "en": [
        "Recheck the official README; do not treat versions in the article as permanent requirements."
      ]
    },
    "posterUrl": "/tutorial-posters/servasyy-local.jpg",
    "tags": [
      "Beginner",
      "ComfyUI",
      "Prompt"
    ],
    "relatedResourceIds": [
      "official-minimax-h3",
      "huangserva-director-workflows"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/servasyy_ai/status/2085251627880255525",
      "author": "huangserva",
      "handle": "@servasyy_ai",
      "publishedAt": "2026-08-06",
      "originalLanguage": "zh"
    },
    "engagement": {
      "replies": 51,
      "reposts": 167,
      "likes": 842,
      "views": 354287,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "原作者长文覆盖安装、模型目录、三种生成模式和 Prompt；本站提供阅读导航，具体操作请进入原文。",
      "en": "The original article covers installation, model folders, three generation modes and prompting. This page is a reading guide; follow the article for the procedure."
    },
    "nextGuideIds": [
      "official-deployment",
      "context-ir-local-pipeline"
    ],
    "guideType": "reference",
    "original": {
      "kind": "x-article",
      "capturedAt": "2026-09-13T09:41:15.626Z"
    }
  },
  {
    "id": "four-essential-h3-repositories",
    "contentType": "community",
    "category": "getting-started",
    "title": {
      "zh": "四个 H3 仓库怎么组合使用",
      "en": "How four essential H3 repositories fit together"
    },
    "outcome": {
      "zh": "把官方 Skill、Turbo、Motion Context 和音频工作流放进一条清晰路线。",
      "en": "Place the official Skill, Turbo, Motion Context, and audio workflows into one clear path."
    },
    "audience": {
      "zh": "已经跑通基础 H3，正在选择扩展能力的用户。",
      "en": "Users who have the base model working and are choosing extensions."
    },
    "hardware": {
      "zh": "取决于选中的本地推理路线。",
      "en": "Depends on the selected local inference route."
    },
    "prerequisites": {
      "zh": [
        "先能运行原生 H3",
        "明确当前瓶颈是 Prompt、速度、时长还是音频"
      ],
      "en": [
        "A working native H3 setup",
        "A clear bottleneck: prompting, speed, duration, or audio"
      ]
    },
    "steps": {
      "zh": [
        "先用官方 Skill 建立 Prompt 基线。",
        "需要速度时单独测试 Turbo。",
        "需要跨段连续性或精细音频时，再分别加入 Motion Context 或 T8。"
      ],
      "en": [
        "Start with the official Skill for a prompt baseline.",
        "Test Turbo alone when speed is the constraint.",
        "Add Motion Context or T8 separately for continuity or detailed audio control."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "优化器并非越多越好；每次只加一个并固定素材对比。"
      ],
      "en": [
        "More optimizers are not automatically better; add one at a time with fixed inputs."
      ]
    },
    "posterUrl": "/tutorial-posters/servasyy-local.jpg",
    "tags": [
      "Roadmap",
      "Turbo",
      "Motion Context",
      "Audio"
    ],
    "relatedResourceIds": [
      "official-minimax-h3",
      "turbo-lora",
      "motion-context",
      "audio-t8"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/servasyy_ai/status/2086101388132835743",
      "author": "huangserva",
      "handle": "@servasyy_ai",
      "publishedAt": "2026-08-08",
      "originalLanguage": "zh"
    },
    "engagement": {
      "replies": 78,
      "reposts": 132,
      "likes": 894,
      "views": 101806,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:41:28.322Z"
    }
  },
  {
    "id": "attention-low-vram-lora",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "Attention、人物 LoRA 与低显存路线补全",
      "en": "Attention backends, character LoRA, and low-VRAM paths"
    },
    "outcome": {
      "zh": "按需求补上更稳参数、Attention 后端、人物 LoRA 或低显存方案。",
      "en": "Add steadier settings, an attention backend, character LoRA, or a low-VRAM path based on the actual need."
    },
    "audience": {
      "zh": "已经有基础工作流，想针对性能或角色一致性继续优化的用户。",
      "en": "Users with a base workflow who want performance or character-consistency improvements."
    },
    "hardware": {
      "zh": "NVIDIA；不同后端与量化组合需单独测试。",
      "en": "NVIDIA; each backend and quantization combination needs separate testing."
    },
    "prerequisites": {
      "zh": [
        "稳定的原生基线",
        "固定 seed、Prompt 与输入素材"
      ],
      "en": [
        "A stable native baseline",
        "A fixed seed, prompt, and input set"
      ]
    },
    "steps": {
      "zh": [
        "先记录原生基线的速度、显存和质量。",
        "每次只切换一个 Attention、LoRA 或量化变量。",
        "保留对比结果，再决定是否进入正式工作流。"
      ],
      "en": [
        "Record baseline speed, VRAM, and quality.",
        "Change one attention, LoRA, or quantization variable at a time.",
        "Keep the comparison before promoting the choice into production."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "原帖是路线总结；具体参数必须回到各项目最新 README 核验。"
      ],
      "en": [
        "The post is a route summary; verify exact parameters in each project's latest README."
      ]
    },
    "posterUrl": "/tutorial-posters/ecosystem.jpg",
    "tags": [
      "Attention",
      "LoRA",
      "Low VRAM"
    ],
    "relatedResourceIds": [
      "spectrum-h3",
      "h3-finetuning",
      "awesome-h3-resources"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/servasyy_ai/status/2087562850781020535",
      "author": "huangserva",
      "handle": "@servasyy_ai",
      "publishedAt": "2026-08-12",
      "originalLanguage": "zh"
    },
    "engagement": {
      "replies": 98,
      "reposts": 125,
      "likes": 526,
      "views": 55196,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:41:30.104Z"
    }
  },
  {
    "id": "context-ir-local-pipeline",
    "contentType": "community",
    "category": "prompt",
    "title": {
      "zh": "H3 Prompt：把镜头、对白与声音写清楚",
      "en": "H3 prompts: plan shots, dialogue and sound"
    },
    "outcome": {
      "zh": "选对生成模式，读懂官方 Prompt 示例，并检查镜头顺序、对白与声音是否对应。",
      "en": "Choose the generation mode, understand an official prompt example, and check the alignment of shots, dialogue and sound."
    },
    "audience": {
      "zh": "已经选好运行方式，想把场景、动作、镜头与声音写清楚的 H3 初学者。",
      "en": "H3 beginners who have chosen a runtime and want to describe scenes, action, camera and sound clearly."
    },
    "hardware": {
      "zh": "阅读不要求本地 GPU；实际生成使用已配置的本地或云端 H3。",
      "en": "No local GPU is needed to read the guide. Rendering needs an existing local or cloud H3 setup."
    },
    "prerequisites": {
      "zh": [
        "明确使用文字、首尾帧还是多模态参考",
        "准备想表达的动作、对白和声音"
      ],
      "en": [
        "Decide between text, boundary frames and multimodal references",
        "Prepare the intended action, dialogue and sound"
      ]
    },
    "steps": {
      "zh": [
        "先按输入选指南：T2VA 与首尾帧用基础指南，Ref2VA 用参考指南；不要把两套结构机械混用。",
        "打开指南中的完整示例，识别 integrated_multimodal_description、overall_soundscape 和 non_diegetic_music 三个部分。",
        "按镜头顺序检查主体动作、机位与切换时间；每次切镜都应有明确的新画面，动作时长要装得进片段。",
        "对白按官方格式指定说话人与语言，环境声和配乐分别放在对应部分；检查有没有无意引入新的音源。",
        "把完整原文示例保存为参考，只修改自己需要的内容；生成后检查动作、台词和声音，再一次只调整一个问题。"
      ],
      "en": [
        "Choose by input: use the base guide for T2VA and boundary frames, and the reference guide for Ref2VA. Do not mechanically mix the structures.",
        "Open a complete example and identify integrated_multimodal_description, overall_soundscape and non_diegetic_music.",
        "Check subject action, camera and cut timing in shot order. Each cut should introduce a clear new view and actions must fit the clip.",
        "Use the official speaker and language format for dialogue. Keep ambience and score in their own sections and check for unintended sound sources.",
        "Save the complete original example as a reference and edit only what you need. Review action, dialogue and sound after generation, then address one issue at a time."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "T2VA / FL2VA 与 Ref2VA 的参考语法不同，应选对应的官方指南。",
        "原作者完整 Prompt 通过来源入口阅读；不要把本站概括当作原文或保证输出一致。"
      ],
      "en": [
        "T2VA / FL2VA and Ref2VA use different reference syntax; use the matching official guide.",
        "Read complete original prompts at the linked source. These summaries are not verbatim prompts or guarantees of identical output."
      ]
    },
    "posterUrl": "/tutorial-posters/context-ir.jpg",
    "tags": [
      "Prompt",
      "Dialogue",
      "Camera"
    ],
    "relatedResourceIds": [
      "official-minimax-h3",
      "minimax-h3-easy"
    ],
    "source": {
      "platform": "huggingface",
      "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md",
      "author": "MiniMax",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "create",
    "depth": "deep",
    "flagship": true,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "official"
    },
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "先看官方完整示例，再理解各字段负责什么；能把镜头、对白与声音安排清楚，比复制一串风格词更可操作。",
      "en": "Start with complete official examples and understand each field. Explicit shots, dialogue and sound give a more useful starting point than a string of style words."
    },
    "learningResources": [
      {
        "label": {
          "zh": "官方基础 Prompt 指南与完整示例",
          "en": "Official base prompt guide and complete examples"
        },
        "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md",
        "kind": "prompt"
      },
      {
        "label": {
          "zh": "参考素材模式 Prompt 指南",
          "en": "Reference-mode prompt guide"
        },
        "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md",
        "kind": "prompt"
      },
      {
        "label": {
          "zh": "观看 Prompt 章节",
          "en": "Watch the prompting chapter"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=1380",
        "kind": "demo"
      }
    ],
    "cost": {
      "zh": "阅读官方指南免费；生成测试使用你的现有平台或硬件。",
      "en": "The official guides are free to read; generation tests use your platform or hardware."
    },
    "applicableVersions": [
      "MiniMax H3 base / reference prompt guides"
    ],
    "difficulty": "beginner",
    "sourceRefs": [
      {
        "title": "Earlier community reference",
        "url": "https://x.com/servasyy_ai/status/2086814254066434076"
      }
    ],
    "hardwareProfiles": [
      "apple-silicon",
      "vram-8",
      "vram-12",
      "vram-16",
      "vram-24-plus",
      "cloud-gpu"
    ],
    "checks": {
      "zh": [
        "Prompt 中引用的素材确实已经提供。",
        "对白原文、说话人和切镜顺序一致。"
      ],
      "en": [
        "Every referenced asset has actually been supplied.",
        "Dialogue text, speakers and shot order are consistent."
      ]
    },
    "expectedResult": {
      "zh": "选对生成模式，读懂官方 Prompt 示例，并检查镜头顺序、对白与声音是否对应。",
      "en": "Choose the generation mode, understand an official prompt example, and check the alignment of shots, dialogue and sound."
    },
    "chapters": [
      {
        "title": {
          "zh": "Prompt 写法",
          "en": "Prompting"
        },
        "url": "https://www.youtube.com/watch?v=G3YHSvXZP_g&t=1380",
        "seconds": 1380
      }
    ],
    "troubleshooting": [
      {
        "problem": {
          "zh": "参考素材被凭空增加",
          "en": "Invented reference assets"
        },
        "solution": {
          "zh": "核对每个 Picture、Video、Audio 标签，只保留实际输入；不要让辅助 LLM 猜测看不到的素材。",
          "en": "Audit every Picture, Video and Audio label and keep only supplied inputs. Do not ask an assistant to guess unseen assets."
        }
      },
      {
        "problem": {
          "zh": "台词与画面不对应",
          "en": "Dialogue does not match the shot"
        },
        "solution": {
          "zh": "按官方示例重新检查说话人标记、语言和镜头时间，缩短过长台词。",
          "en": "Recheck speaker tags, language and shot timing against the official example, and shorten dialogue that does not fit."
        }
      }
    ],
    "nextGuideIds": [
      "native-comfyui-modes",
      "video-pov-camera-angles"
    ],
    "relatedCases": [
      {
        "id": "official-t2va-starship",
        "title": {
          "zh": "舰桥上的跃迁余震：官方完整 Prompt",
          "en": "After the Fleet Jumps: complete official prompt"
        },
        "relationship": "example"
      }
    ],
    "learningOrder": 1,
    "guideType": "project",
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T09:41:32.429Z"
    }
  },
  {
    "id": "video-pov-camera-angles",
    "contentType": "community",
    "category": "prompt",
    "title": {
      "zh": "人物参考与对白：分配图像、视频和声音",
      "en": "Character references and dialogue: assign image, video and audio roles"
    },
    "outcome": {
      "zh": "用 Ref2VA 指定人物与声音参考，检查生成对白及身份保持；机位编辑作为后续变化。",
      "en": "Use Ref2VA to assign identity and sound references, then review dialogue and identity retention before trying camera edits."
    },
    "audience": {
      "zh": "已能生成单段 H3 视频，希望用参考素材约束人物特征并组织对白的创作者。",
      "en": "Creators with a working single-clip H3 setup who want reference media to guide identity and dialogue."
    },
    "hardware": {
      "zh": "已能运行 H3 的 ComfyUI；需另备 Ref2VA 权重，输入越复杂计算负担越高。",
      "en": "ComfyUI already running H3, with separate Ref2VA weights. More complex inputs increase compute requirements."
    },
    "prerequisites": {
      "zh": [
        "可使用的人物图或视频，以及需要的声音参考",
        "知道哪些内容要保留、哪些要改变"
      ],
      "en": [
        "A usable character image or video and any required sound references",
        "A clear decision on what to retain and what to change"
      ]
    },
    "steps": {
      "zh": [
        "打开 R2V 模板，确认模型切换到 Ref2VA；不要继续使用 T2V/I2V 的 FL2VA 权重。",
        "按模板连接参考图、视频和音频，记下输入顺序；第一次只保留任务必需的少量素材。",
        "对照官方参考指南声明各素材的任务：身份、动作、背景或声音；明确保留和修改的部分。",
        "在同一短场景里安排一句对白，按官方格式标出说话人和语言；检查所有引用都对应已上传素材。",
        "生成后分别检查人物特征、动作、台词与声音。人物不稳先减少互相冲突的参考，台词不完整先缩短对白。",
        "基线满意后再尝试改变机位或动作，并写清哪些内容保持不变；不要同时替换所有参考。"
      ],
      "en": [
        "Open the R2V template and select Ref2VA instead of the FL2VA weights used for T2V/I2V.",
        "Connect image, video and audio references as shown in the template and record their input order. Start with only the assets the task requires.",
        "Use the reference guide to assign identity, action, background or sound roles and specify what should be retained or changed.",
        "Place one spoken line in a short scene, with the official speaker and language format. Check that every reference corresponds to an uploaded input.",
        "Inspect identity, motion, dialogue and sound separately. Reduce conflicting references for unstable identity and shorten incomplete dialogue.",
        "Once the baseline is satisfactory, try a camera or action change while specifying what stays the same. Do not replace all references at once."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "不要声称像素级复刻；快速运动和遮挡仍可能漂移。"
      ],
      "en": [
        "Do not expect pixel-level reconstruction; fast motion and occlusion can still drift."
      ]
    },
    "posterUrl": "/tutorial-posters/brand-prompt.svg",
    "tags": [
      "Ref2VA",
      "References",
      "Dialogue"
    ],
    "relatedResourceIds": [
      "official-minimax-h3",
      "aimixer-h3-director"
    ],
    "source": {
      "platform": "docs",
      "url": "https://docs.comfy.org/tutorials/video/minimax/minimax-h3-native#minimax-h3-reference-to-video-r2v",
      "author": "ComfyUI",
      "authorProfileUrl": "https://github.com/Comfy-Org",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "create",
    "depth": "deep",
    "flagship": true,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "official"
    },
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "从“每个参考负责什么”入手，区分人物、动作、机位和声音来源，避免把素材堆进去却没有明确任务。",
      "en": "Assign a role to each reference: identity, motion, camera or sound. The guide makes the intended use of each input explicit."
    },
    "learningResources": [
      {
        "label": {
          "zh": "官方 R2V 工作流 JSON",
          "en": "Official R2V workflow JSON"
        },
        "url": "https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json",
        "kind": "workflow"
      },
      {
        "label": {
          "zh": "参考模式完整 Prompt 与素材示例",
          "en": "Complete reference prompts and input examples"
        },
        "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md",
        "kind": "prompt"
      },
      {
        "label": {
          "zh": "Ref2VA 模型文件",
          "en": "Ref2VA model files"
        },
        "url": "https://huggingface.co/Comfy-Org/MiniMax-H3",
        "kind": "models"
      }
    ],
    "cost": {
      "zh": "本地生成占用自己的硬件、内存和磁盘；模型与工具的使用条件请看原始说明。",
      "en": "Local generation uses your own hardware, RAM and storage. See the original model and tool terms."
    },
    "applicableVersions": [
      "ComfyUI 0.30+ · native H3 templates"
    ],
    "difficulty": "intermediate",
    "sourceRefs": [
      {
        "title": "Earlier community reference",
        "url": "https://x.com/magnific/status/2090423731055448271"
      }
    ],
    "hardwareProfiles": [
      "vram-16",
      "vram-24-plus",
      "cloud-gpu"
    ],
    "checks": {
      "zh": [
        "模型与参考模式一致，引用次序正确。",
        "人物、台词和声音分别完成检查；不把“有声音”当作“对白正确”。"
      ],
      "en": [
        "The model matches reference mode and references are ordered correctly.",
        "Identity, dialogue and sound have each been checked; audible output is not proof of correct dialogue."
      ]
    },
    "expectedResult": {
      "zh": "用 Ref2VA 指定人物与声音参考，检查生成对白及身份保持；机位编辑作为后续变化。",
      "en": "Use Ref2VA to assign identity and sound references, then review dialogue and identity retention before trying camera edits."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "人物漂移",
          "en": "Identity drift"
        },
        "solution": {
          "zh": "减少冲突参考，明确哪个素材提供身份、哪个提供动作；保留同一短场景做对比。",
          "en": "Reduce conflicting inputs and separate identity from action references. Compare within the same short scene."
        }
      },
      {
        "problem": {
          "zh": "对白被截断或声音不符合预期",
          "en": "Truncated dialogue or unexpected sound"
        },
        "solution": {
          "zh": "缩短台词，检查语言与说话人标记，以及音频是引用还是复用。",
          "en": "Shorten the line and check language, speaker tags and whether audio is referenced or reused."
        }
      }
    ],
    "nextGuideIds": [
      "context-ir-local-pipeline",
      "prompt-agent-skill"
    ],
    "relatedCases": [
      {
        "id": "official-ref2va-lamb",
        "title": {
          "zh": "粉色西装与黑色羔羊：官方参考与对白示例",
          "en": "Follow the Wind: official reference and dialogue example"
        },
        "relationship": "example"
      }
    ],
    "learningOrder": 3,
    "guideType": "project",
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T12:44:41.834Z"
    }
  },
  {
    "id": "storyboard-to-trailer",
    "contentType": "community",
    "category": "prompt",
    "title": {
      "zh": "从分镜关键帧生成风格统一的预告片",
      "en": "Turn storyboard frames into a consistent trailer"
    },
    "outcome": {
      "zh": "用最终影片关键帧与镜头描述逐镜生成，再组合成预告片。",
      "en": "Generate shot by shot from final-film frames and shot descriptions, then assemble a trailer."
    },
    "audience": {
      "zh": "已有分镜或成片关键帧的电影与广告创作者。",
      "en": "Film and advertising creators with storyboards or final-film frames."
    },
    "hardware": {
      "zh": "支持图像参考的 H3 工作流。",
      "en": "An H3 image-reference workflow."
    },
    "prerequisites": {
      "zh": [
        "按镜头切出的参考帧",
        "每个镜头的动作与运镜描述",
        "剪辑时间线"
      ],
      "en": [
        "Reference frames split by shot",
        "Action and camera direction for each shot",
        "An editing timeline"
      ]
    },
    "steps": {
      "zh": [
        "从最终视觉中提取每个镜头的参考帧。",
        "逐镜输入参考帧与镜头描述，固定风格约束。",
        "筛选结果并按节奏剪辑，最后统一声音与色彩。"
      ],
      "en": [
        "Extract one reference frame for every intended shot.",
        "Generate each shot with its frame and description while keeping style constraints fixed.",
        "Select takes, edit to rhythm, and finish sound and color consistently."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "镜头间一致性仍需人工挑片与后期，不应一次生成整条预告。"
      ],
      "en": [
        "Cross-shot consistency still needs curation and post; do not expect one-pass trailer generation."
      ]
    },
    "posterUrl": "/tutorial-posters/storyboard.jpg",
    "tags": [
      "Storyboard",
      "Trailer",
      "Consistency"
    ],
    "relatedResourceIds": [
      "aimixer-h3-director",
      "huangserva-director-workflows",
      "openmontage"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/magnific/status/2090508187875356957",
      "author": "Magnific",
      "handle": "@magnific",
      "publishedAt": "2026-08-21",
      "originalLanguage": "en"
    },
    "engagement": {
      "replies": 7,
      "reposts": 10,
      "likes": 141,
      "views": 11000,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "create",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "project",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:41:46.689Z"
    }
  },
  {
    "id": "3060-turbo-trailer",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "RTX 3060 + Turbo LoRA 制作一分钟预告片",
      "en": "Build a one-minute trailer on RTX 3060 with Turbo LoRA"
    },
    "outcome": {
      "zh": "把多张既有图片转成多个短镜头，再剪成一分钟预告片。",
      "en": "Animate an existing image set into short shots and edit them into a one-minute trailer."
    },
    "audience": {
      "zh": "12GB 显存、已有概念图或分镜图的创作者。",
      "en": "Creators with 12GB VRAM and an existing concept-art or storyboard set."
    },
    "hardware": {
      "zh": "RTX 3060 12GB、24GB 系统内存（原帖实测环境）。",
      "en": "RTX 3060 12GB and 24GB system memory, as reported by the source."
    },
    "prerequisites": {
      "zh": [
        "一组风格统一的图片",
        "可运行的 Turbo LoRA 工作流",
        "视频剪辑工具"
      ],
      "en": [
        "A style-consistent image set",
        "A working Turbo LoRA workflow",
        "A video editor"
      ]
    },
    "steps": {
      "zh": [
        "按镜头顺序整理图片与动作描述。",
        "用 Turbo LoRA 逐张生成短片，不在单次生成中追求整分钟。",
        "挑选结果并在剪辑软件中完成节奏与声音。"
      ],
      "en": [
        "Order the images and action notes by shot.",
        "Generate short clips with Turbo LoRA instead of chasing a one-minute single pass.",
        "Select takes and finish pacing and sound in an editor."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "一分钟是后期拼接结果，不是 H3 单次生成时长。"
      ],
      "en": [
        "The one-minute result is edited from shots; it is not a single H3 generation."
      ]
    },
    "posterUrl": "/tutorial-posters/3060-trailer.jpg",
    "tags": [
      "RTX 3060",
      "Turbo LoRA",
      "Trailer"
    ],
    "relatedResourceIds": [
      "turbo-lora",
      "aimixer-h3-director"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/PixelAigc/status/2086958485741441240",
      "author": "PixelAIGC",
      "handle": "@PixelAigc",
      "publishedAt": "2026-08-11",
      "originalLanguage": "zh"
    },
    "engagement": {
      "replies": 30,
      "reposts": 25,
      "likes": 261,
      "views": 22584,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "project",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:42:13.897Z"
    }
  },
  {
    "id": "maestro-six-step-turbo",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "Maestro 六步 Turbo：先以速度换预览",
      "en": "Maestro six-step Turbo: trade detail for iteration speed"
    },
    "outcome": {
      "zh": "在 Maestro 中用六步蒸馏 LoRA 快速预览 H3 结果。",
      "en": "Use a six-step distilled LoRA in Maestro for fast H3 previews."
    },
    "audience": {
      "zh": "使用 Maestro，希望缩短试错时间的本地用户。",
      "en": "Local Maestro users who want shorter iteration cycles."
    },
    "hardware": {
      "zh": "取决于 Maestro 与所选 H3 权重。",
      "en": "Depends on Maestro and the selected H3 weights."
    },
    "prerequisites": {
      "zh": [
        "可运行的 Maestro H3 工作流",
        "Turbo / Distilled LoRA",
        "固定基线素材"
      ],
      "en": [
        "A working Maestro H3 workflow",
        "A Turbo or distilled LoRA",
        "A fixed baseline input"
      ]
    },
    "steps": {
      "zh": [
        "先在原生模式保存基线。",
        "切到六步 Turbo 并保持其他参数不变。",
        "把 Turbo 用于选镜和构图，重要成片再比较更高步数。"
      ],
      "en": [
        "Save a native-mode baseline first.",
        "Switch to six-step Turbo without changing other variables.",
        "Use Turbo for shot selection and compare higher steps for important finals."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "原作者明确提醒结果仍可能粗糙；速度提升不等于同等质量。"
      ],
      "en": [
        "The source explicitly notes rougher results; faster does not mean equivalent quality."
      ]
    },
    "posterUrl": "/tutorial-posters/maestro-six-step.jpg",
    "tags": [
      "Maestro",
      "6 Steps",
      "Turbo"
    ],
    "relatedResourceIds": [
      "turbo-lora",
      "modeltc-h3-turbo"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/blizaine/status/2085554573494657328",
      "author": "Blaine Brown",
      "handle": "@blizaine",
      "publishedAt": "2026-08-07",
      "originalLanguage": "en"
    },
    "engagement": {
      "replies": 9,
      "reposts": 7,
      "likes": 103,
      "views": 27115,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:43:59.398Z"
    }
  },
  {
    "id": "maestro-pruned-20b",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "Maestro 低显存：Pruned 20B + Turbo",
      "en": "Low-VRAM Maestro: pruned 20B plus Turbo"
    },
    "outcome": {
      "zh": "用剪枝 20B、六步 Turbo 与可选 First Block Cache 降低显存压力。",
      "en": "Reduce VRAM pressure with the pruned 20B model, six-step Turbo, and optional First Block Cache."
    },
    "audience": {
      "zh": "显存不足以舒适运行全量模型的 Maestro 用户。",
      "en": "Maestro users who cannot comfortably run the full model."
    },
    "hardware": {
      "zh": "低显存 NVIDIA 环境；实际下限以 Maestro 当前文档为准。",
      "en": "A lower-VRAM NVIDIA setup; use current Maestro documentation for the actual minimum."
    },
    "prerequisites": {
      "zh": [
        "Maestro 新版",
        "Pruned 20B 权重",
        "Turbo 权重"
      ],
      "en": [
        "A current Maestro release",
        "Pruned 20B weights",
        "Turbo weights"
      ]
    },
    "steps": {
      "zh": [
        "先选择 Pruned 20B 与六步 Turbo。",
        "完成固定素材基线并记录显存。",
        "仍需优化时再单独打开 First Block Cache 并检查画质。"
      ],
      "en": [
        "Select the pruned 20B model and six-step Turbo.",
        "Run a fixed-input baseline and record VRAM.",
        "Enable First Block Cache only if still needed, then inspect quality again."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "Cache 是额外变量，不能和 Turbo 一起直接宣称无损。"
      ],
      "en": [
        "Cache is another variable; do not claim lossless stacking with Turbo without testing."
      ]
    },
    "posterUrl": "/tutorial-posters/maestro-low-vram.jpg",
    "tags": [
      "Low VRAM",
      "Pruned 20B",
      "Cache"
    ],
    "relatedResourceIds": [
      "turbo-lora",
      "awesome-h3-resources"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/cocktailpeanut/status/2086119385757041048",
      "author": "cocktail peanut",
      "handle": "@cocktailpeanut",
      "publishedAt": "2026-08-08",
      "originalLanguage": "en"
    },
    "engagement": {
      "replies": 10,
      "reposts": 11,
      "likes": 130,
      "views": 12486,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:44:17.581Z"
    }
  },
  {
    "id": "3060-turbo-benchmark",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "RTX 3060 的 Turbo LoRA 前后实测",
      "en": "Before-and-after Turbo LoRA benchmark on RTX 3060"
    },
    "outcome": {
      "zh": "在同一台 3060 上用固定分辨率和时长比较原生与 Turbo。",
      "en": "Compare native and Turbo runs at fixed resolution and duration on the same RTX 3060."
    },
    "audience": {
      "zh": "想用真实生成时间评估 Turbo 收益的 12GB 用户。",
      "en": "12GB users who want a measured view of Turbo gains."
    },
    "hardware": {
      "zh": "RTX 3060 12GB、64GB RAM、Windows 11（原帖实测）。",
      "en": "RTX 3060 12GB, 64GB RAM, and Windows 11, as reported by the source."
    },
    "prerequisites": {
      "zh": [
        "1056×608、10 秒固定测试",
        "v4 step600 EMA Turbo LoRA"
      ],
      "en": [
        "A fixed 1056×608 ten-second test",
        "The v4 step600 EMA Turbo LoRA"
      ]
    },
    "steps": {
      "zh": [
        "先跑原生基线并记录时间。",
        "只加入指定 Turbo LoRA，保持素材与输出设置不变。",
        "比较耗时、画面、声音与运动稳定性。"
      ],
      "en": [
        "Run a native baseline and record its time.",
        "Add only the specified Turbo LoRA while keeping input and output settings fixed.",
        "Compare time, image quality, audio, and motion stability."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "原帖时间只代表该机器与设置，不能外推到所有环境。"
      ],
      "en": [
        "The reported times apply only to that machine and configuration."
      ]
    },
    "posterUrl": "/tutorial-posters/aipix-turbo.jpg",
    "tags": [
      "Benchmark",
      "RTX 3060",
      "Turbo"
    ],
    "relatedResourceIds": [
      "turbo-lora"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/AIPixLab/status/2088797656118390866",
      "author": "AI Photo Lab",
      "handle": "@AIPixLab",
      "publishedAt": "2026-08-16",
      "originalLanguage": "ja"
    },
    "engagement": {
      "replies": 1,
      "reposts": 1,
      "likes": 48,
      "views": 11943,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:44:31.622Z"
    }
  },
  {
    "id": "mac-phosphene-turbo",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "Mac Studio 上用 Phosphene 跑 Turbo",
      "en": "Run H3 Turbo on Mac Studio with Phosphene"
    },
    "outcome": {
      "zh": "在 Apple Silicon 上用较少步数把 5 秒生成从约 20 分钟压到约 8 分钟。",
      "en": "Use fewer steps on Apple Silicon to reduce a five-second run from about 20 minutes to about eight in the reported setup."
    },
    "audience": {
      "zh": "希望用图形化工具体验 Mac H3 Turbo 的用户。",
      "en": "Mac users who want a graphical H3 Turbo path."
    },
    "hardware": {
      "zh": "Mac Studio M4 64GB（原帖实测）。",
      "en": "Mac Studio M4 with 64GB memory, as reported by the source."
    },
    "prerequisites": {
      "zh": [
        "支持 H3 的 Phosphene 版本",
        "Turbo 模式权重",
        "短片固定测试"
      ],
      "en": [
        "A Phosphene version with H3 support",
        "Turbo-mode weights",
        "A fixed short-clip test"
      ]
    },
    "steps": {
      "zh": [
        "先在关闭 Turbo 时生成基线。",
        "打开 Turbo 并保持 Prompt 与素材不变。",
        "把快速结果用于迭代，重要镜头再回到高质量设置。"
      ],
      "en": [
        "Generate a baseline with Turbo disabled.",
        "Enable Turbo while keeping prompt and media unchanged.",
        "Use the faster result for iteration and return to quality settings for important shots."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "原作者报告 Turbo 运动和形变仍较弱，关闭 Turbo 质量更好。"
      ],
      "en": [
        "The source reports weaker motion and distortion under Turbo, with better quality when disabled."
      ]
    },
    "posterUrl": "/tutorial-posters/mac-turbo.jpg",
    "tags": [
      "Mac",
      "Phosphene",
      "Turbo"
    ],
    "relatedResourceIds": [
      "antirez-h3-metal",
      "turbo-lora"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/cocktailpeanut/status/2085765919079350355",
      "author": "cocktail peanut",
      "handle": "@cocktailpeanut",
      "publishedAt": "2026-08-08",
      "originalLanguage": "en"
    },
    "engagement": {
      "replies": 4,
      "reposts": 11,
      "likes": 76,
      "views": 9875,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:44:46.946Z"
    }
  },
  {
    "id": "rent-gpu-comfyui",
    "contentType": "community",
    "category": "getting-started",
    "title": {
      "zh": "云端免部署：完成第一条 H3 视频",
      "en": "H3 in the cloud: your first video without local setup"
    },
    "outcome": {
      "zh": "从云端模板开始，替换 Prompt 或首帧，生成并下载一条带声 MP4。",
      "en": "Open a cloud template, replace its prompt or first frame, then generate and download an MP4 with audio."
    },
    "audience": {
      "zh": "本地显存不足、只想先验证效果的用户。",
      "en": "Users with limited local VRAM who want to validate results first."
    },
    "hardware": {
      "zh": "能打开 Comfy Cloud 的浏览器与稳定网络；本机无需 NVIDIA GPU。",
      "en": "A browser that can access Comfy Cloud and a stable connection; no local NVIDIA GPU is required."
    },
    "prerequisites": {
      "zh": [
        "可使用 Comfy Cloud 的账户与足够额度",
        "一段场景描述；图生视频需自备参考图"
      ],
      "en": [
        "A Comfy Cloud account with sufficient credits",
        "A scene description; supply an image for image-to-video"
      ]
    },
    "steps": {
      "zh": [
        "打开上方官方云端模板，完成账户登录；先检查当前计划、可用额度和生成入口。",
        "从 MiniMax H3 T2V 模板开始；图生视频则打开 I2V 模板，并在 Load Image 中上传自己的首帧。",
        "保留模板模型与预览尺寸，替换场景、主体动作和声音描述；确认图生视频的首帧输入已连接。",
        "核对时长、比例与额度提示后运行一次；在任务完成前不要重复提交同一请求。",
        "在输出节点播放结果，检查画面、声音和时长，再使用输出的下载入口保存 MP4；保存工作流方便下次修改。"
      ],
      "en": [
        "Open the official cloud template and sign in. Check the account plan, credits and generation controls.",
        "Start with MiniMax H3 T2V. For I2V, open its template and upload your first frame through Load Image.",
        "Keep the template models and preview size. Replace the scene, action and sound description; check the first-frame input for I2V.",
        "Check duration, aspect ratio and credit information, then run once. Do not resubmit the same request while it is pending.",
        "Play the output and check picture, sound and duration. Download the MP4 from the output and save the workflow for later changes."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "云端可用模型与节点随服务更新；原作中的本地路径不适用于云端。"
      ],
      "en": [
        "Cloud models and nodes can change with service updates. Local filesystem paths from other guides do not apply here."
      ]
    },
    "posterUrl": "/tutorial-posters/comfy-cloud-docs.png",
    "tags": [
      "Cloud GPU",
      "ComfyUI",
      "Intermediate"
    ],
    "relatedResourceIds": [
      "official-minimax-h3"
    ],
    "source": {
      "platform": "docs",
      "url": "https://docs.comfy.org/get_started/cloud",
      "author": "ComfyUI",
      "authorProfileUrl": "https://github.com/Comfy-Org",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "deep",
    "flagship": true,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "official"
    },
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "不具备合适本地 GPU 时，从官方云端模板入手，省去本机模型安装；生成前先看账户额度。",
      "en": "Use the official cloud template when local hardware is unsuitable. It avoids local model installation; check account credits before rendering."
    },
    "learningResources": [
      {
        "label": {
          "zh": "在 Comfy Cloud 打开 T2V 模板",
          "en": "Open the T2V template in Comfy Cloud"
        },
        "url": "https://cloud.comfy.org/?template=video_minimax_h3_t2v",
        "kind": "workflow"
      },
      {
        "label": {
          "zh": "Comfy Cloud 入门说明",
          "en": "Comfy Cloud onboarding"
        },
        "url": "https://docs.comfy.org/get_started/cloud",
        "kind": "documentation"
      },
      {
        "label": {
          "zh": "图生视频云端模板",
          "en": "Cloud image-to-video template"
        },
        "url": "https://cloud.comfy.org/?template=video_minimax_h3_i2v",
        "kind": "workflow"
      }
    ],
    "cost": {
      "zh": "Comfy Cloud 使用账户计划与额度，生成会消耗额度。实际费用以运行前页面为准。",
      "en": "Comfy Cloud uses an account plan and credits. Rendering consumes credits; check the current page before running."
    },
    "applicableVersions": [
      "Comfy Cloud · native MiniMax H3 templates"
    ],
    "difficulty": "intermediate",
    "sourceRefs": [
      {
        "title": "Earlier community reference",
        "url": "https://x.com/Lonely__MH/status/2085302359971053826"
      }
    ],
    "hardwareProfiles": [
      "cloud-gpu"
    ],
    "checks": {
      "zh": [
        "下载的 MP4 能在普通播放器中播放画面和声音。",
        "记录本次实际消耗；不要把模板入口免费等同于生成免费。"
      ],
      "en": [
        "The downloaded MP4 plays with picture and sound in a regular player.",
        "Record the actual credit usage; an accessible template does not mean free rendering."
      ]
    },
    "expectedResult": {
      "zh": "从云端模板开始，替换 Prompt 或首帧，生成并下载一条带声 MP4。",
      "en": "Open a cloud template, replace its prompt or first frame, then generate and download an MP4 with audio."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "账户或额度无法运行",
          "en": "Account or credit restriction"
        },
        "solution": {
          "zh": "按云端账户页面处理计划或额度问题；不要改本机驱动来解决云端计费问题。",
          "en": "Resolve the plan or credit issue in the cloud account page; local driver changes cannot fix cloud billing."
        }
      },
      {
        "problem": {
          "zh": "输入图没有参与生成",
          "en": "The input image has no effect"
        },
        "solution": {
          "zh": "确认打开的是 I2V 模板且图片已上传到对应输入。",
          "en": "Confirm that the I2V template is open and the image is uploaded to its input."
        }
      }
    ],
    "nextGuideIds": [
      "native-comfyui-modes",
      "context-ir-local-pipeline"
    ],
    "learningOrder": 3,
    "guideType": "setup",
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T09:44:49.334Z"
    }
  },
  {
    "id": "dual-clock-audio",
    "contentType": "community",
    "category": "audio",
    "title": {
      "zh": "双时钟采样器：分别控制视频与音频步数",
      "en": "Dual-clock sampling: separate video and audio step counts"
    },
    "outcome": {
      "zh": "在 ComfyUI 中分别控制音画采样，并与加速 LoRA 做受控对比。",
      "en": "Control video and audio sampling separately in ComfyUI and compare it carefully with an acceleration LoRA."
    },
    "audience": {
      "zh": "对声音质量、同步和生成速度都敏感的进阶用户。",
      "en": "Advanced users balancing sound quality, sync, and generation speed."
    },
    "hardware": {
      "zh": "能运行 H3 ComfyUI 的 NVIDIA 环境。",
      "en": "An NVIDIA environment capable of running H3 in ComfyUI."
    },
    "prerequisites": {
      "zh": [
        "T8 音画节点",
        "官方 H3 工作流",
        "固定音画测试 Prompt"
      ],
      "en": [
        "The T8 audio-video nodes",
        "A native H3 workflow",
        "A fixed audio-video test prompt"
      ]
    },
    "steps": {
      "zh": [
        "先确认官方 H3 ComfyUI 基线可以输出带声 MP4，并保存固定 seed、Prompt、画布、帧数和 20 步结果。",
        "在 ComfyUI Manager 搜索 MiniMax H3 Audio T8，或用 comfy CLI 安装；完成后彻底重启 ComfyUI。",
        "打开 examples/workflows/01-basic-generation，替换成自己的模型与素材；首轮使用 22 帧、小画布和 stock sampler。",
        "先接 MiniMaxH3AudioConditioningT8、MiniMaxH3DualClockSamplerT8、MiniMaxH3AVDecodeT8 和 MiniMaxH3PreflightT8，运行预检并修复尺寸、标签或显存错误。",
        "固定视频步数，只改变音频时钟；再固定音频步数，只改变视频时钟。每次记录总耗时、声音清晰度、事件同步和输出轨道信息。",
        "只有稳定链路通过后才加入 Turbo LoRA；实验节点、SageAttention 和长视频链路分开验证，不在第一轮叠加。"
      ],
      "en": [
        "Confirm the official H3 ComfyUI baseline outputs an MP4 with audio, then save its seed, prompt, canvas, frame count, and 20-step result.",
        "Find MiniMax H3 Audio T8 in ComfyUI Manager or install it with the comfy CLI, then fully restart ComfyUI.",
        "Open examples/workflows/01-basic-generation and replace the models and media. Start with 22 frames, a small canvas, and the stock sampler.",
        "Wire MiniMaxH3AudioConditioningT8, MiniMaxH3DualClockSamplerT8, MiniMaxH3AVDecodeT8, and MiniMaxH3PreflightT8 first. Run preflight and fix size, tag, or VRAM errors.",
        "Hold video steps constant while changing the audio clock, then reverse the test. Record total time, speech clarity, event sync, and output stream metadata each time.",
        "Add Turbo LoRA only after the stable path passes. Test experimental nodes, SageAttention, and long-video graphs separately rather than stacking them in the first run."
      ]
    },
    "commands": [
      "comfy node install minimax-h3-audio-t8",
      "cd ComfyUI/custom_nodes && git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8"
    ],
    "checks": {
      "zh": [
        "启动日志显示插件成功注册且没有导入错误",
        "Preflight 没有模型、尺寸、标签或显存阻塞",
        "基础双时钟工作流输出含视频轨和 32kHz 立体声音频轨",
        "已保存仅改变一个时钟参数的 A/B 结果"
      ],
      "en": [
        "Startup logs show the plugin registered without import errors",
        "Preflight reports no blocking model, size, tag, or VRAM errors",
        "The base dual-clock graph outputs video plus a 32kHz stereo audio stream",
        "A/B results change only one clock parameter at a time"
      ]
    },
    "caveats": {
      "zh": [
        "社区宣称的速度提升需在自己的机器上复测；实验节点与稳定节点要分开。"
      ],
      "en": [
        "Re-test community speed claims on your own hardware and keep experimental nodes separate from stable ones."
      ]
    },
    "posterUrl": "/tutorial-posters/dual-clock.jpg",
    "tags": [
      "Audio",
      "Dual Clock",
      "ComfyUI"
    ],
    "relatedResourceIds": [
      "audio-t8",
      "turbo-lora"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/0xmomonifty/status/2085377094461333550",
      "author": "0xMomo",
      "handle": "@0xmomonifty",
      "publishedAt": "2026-08-06",
      "originalLanguage": "zh"
    },
    "engagement": {
      "replies": 25,
      "reposts": 61,
      "likes": 443,
      "views": 57000,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": false,
    "difficulty": "advanced",
    "estimatedMinutes": 75,
    "hardwareProfiles": [
      "vram-16",
      "vram-24-plus",
      "cloud-gpu"
    ],
    "expectedResult": {
      "zh": "得到一组只改变视频或音频时钟的带声 MP4 对照，并能用 Preflight 报告说明所用模型、帧数、尺寸与显存风险。",
      "en": "Produce an MP4 comparison that changes only the video or audio clock, with a Preflight report documenting models, frames, dimensions, and VRAM risk."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "prompt media tag validation failed",
          "en": "prompt media tag validation failed"
        },
        "solution": {
          "zh": "检查 <Picture 1>、<Video 1>、<Audio 1> 的编号与实际接入顺序，多个参考输入不得省略编号。",
          "en": "Match <Picture 1>, <Video 1>, and <Audio 1> numbering to the wired inputs; do not omit indices with multiple references."
        }
      },
      {
        "problem": {
          "zh": "OOM、NaN 或音频丢失",
          "en": "OOM, NaN, or missing audio"
        },
        "solution": {
          "zh": "保存完整 Error Report，从 22 帧小画布和稳定节点重跑，并清理实验 patch 与缓存。",
          "en": "Save the complete Error Report, retry at 22 frames on a small canvas with stable nodes, and remove experimental patches and stale caches."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "删除 ComfyUI/custom_nodes/minimax-h3-audio-T8 后重启。",
        "可选提示词重写器依赖只在使用该功能时安装，不要为基础音频工作流全局升级 Transformers。"
      ],
      "en": [
        "Delete ComfyUI/custom_nodes/minimax-h3-audio-T8 and restart.",
        "Install optional prompt-rewriter dependencies only when using that feature; do not globally upgrade Transformers for the base audio graph."
      ]
    },
    "sourceRefs": [
      {
        "title": "MiniMax H3 Audio T8",
        "url": "https://github.com/T8mars/comfyui-minimax-h3-audio-T8"
      }
    ],
    "learningTrack": "create",
    "depth": "guide",
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "applicableVersions": [
      "MiniMax H3 Audio T8 1.44.0",
      "ComfyUI native H3"
    ],
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "project",
    "commandItems": [
      {
        "kind": "command",
        "value": "comfy node install minimax-h3-audio-t8",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "cd ComfyUI/custom_nodes && git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8",
        "platform": "shell"
      }
    ],
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:44:58.878Z"
    }
  },
  {
    "id": "native-comfyui-modes",
    "contentType": "community",
    "category": "comfyui",
    "title": {
      "zh": "图生视频：首帧、尾帧与模式选择",
      "en": "Image to video: first frames, last frames and modes"
    },
    "outcome": {
      "zh": "让一张图片动起来，或用首尾两张图约束开场与结尾。",
      "en": "Animate a still image or constrain the opening and ending with two images."
    },
    "audience": {
      "zh": "第一次在 ComfyUI 中看到 H3 多种工作流的用户。",
      "en": "Users seeing H3's multiple ComfyUI workflows for the first time."
    },
    "hardware": {
      "zh": "已能运行 H3 的本地或云端 ComfyUI；图生视频复用 FL2VA 权重。",
      "en": "Local or cloud ComfyUI already running H3; image-to-video reuses FL2VA weights."
    },
    "prerequisites": {
      "zh": [
        "一张首帧图；需要指定结尾时再准备尾帧图",
        "先通过入门教程确认模型和两个 VAE 可用"
      ],
      "en": [
        "One first-frame image; add a last-frame image only when the ending must be constrained",
        "Complete the first-run guide so the model and both VAEs are available"
      ]
    },
    "steps": {
      "zh": [
        "打开官方 I2V JSON 或模板库中的 MiniMax H3 I2V，确认加载 FL2VA 模型。",
        "上传首帧并连接 first_frame；只想让一张图动起来时，保留 last_frame 为空。",
        "需要首尾帧控制时，再上传尾图并连接 last_frame；两张图应能通过一个清晰动作连接。",
        "设置与构图相符的宽高比，先用模板预览尺寸；参照官方例子写从输入图开始如何运动。",
        "运行后检查开头是否符合首图、结尾是否接近尾图、主体有没有不必要变化；一次调整构图或动作之一。"
      ],
      "en": [
        "Open the official I2V JSON or MiniMax H3 I2V template and confirm the FL2VA model.",
        "Upload the opening image and connect first_frame. Leave last_frame empty when animating a single still.",
        "For boundary-frame control, upload the ending image to last_frame. Choose two images that a clear action can connect.",
        "Choose an aspect ratio that suits the composition and begin at preview size. Follow the official example to describe motion starting from the input.",
        "After rendering, compare the opening and ending with the inputs and inspect unwanted subject changes. Adjust either composition or action at a time."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "不同输入模式对应的权重和接线不同，不要仅替换输入端口。"
      ],
      "en": [
        "Input modes use different weights and wiring; do not merely swap input ports."
      ]
    },
    "posterUrl": "/tutorial-posters/comfyui-native.jpg",
    "tags": [
      "FL2VA",
      "Image to Video",
      "First / Last Frame"
    ],
    "relatedResourceIds": [
      "official-minimax-h3",
      "minimax-h3-easy"
    ],
    "source": {
      "platform": "docs",
      "url": "https://docs.comfy.org/tutorials/video/minimax/minimax-h3-native#minimax-h3-image-to-video-i2v",
      "author": "ComfyUI",
      "authorProfileUrl": "https://github.com/Comfy-Org",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-06",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "create",
    "depth": "deep",
    "flagship": true,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-09-06",
      "basis": "official"
    },
    "materialsNote": {
      "zh": "链接提供原作者工作流、模型或示例；自选参考素材需自行准备。本站没有另行实测这套生成流程。",
      "en": "Links lead to the original workflows, models or examples. Supply your own reference assets where needed. This generation workflow has not been independently run by this site."
    },
    "recommendation": {
      "zh": "把“首帧、尾帧和参考图”分清楚，再用原生模板连接输入，适合从静态图片开始创作。",
      "en": "Distinguish first frames, last frames and general references, then connect inputs in the native template. A practical route from a still image."
    },
    "learningResources": [
      {
        "label": {
          "zh": "官方 I2V 工作流 JSON",
          "en": "Official I2V workflow JSON"
        },
        "url": "https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_i2v.json",
        "kind": "workflow"
      },
      {
        "label": {
          "zh": "基础 Prompt 与首尾帧示例",
          "en": "Base prompts and boundary-frame examples"
        },
        "url": "https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md",
        "kind": "prompt"
      },
      {
        "label": {
          "zh": "模型文件",
          "en": "Model files"
        },
        "url": "https://huggingface.co/Comfy-Org/MiniMax-H3",
        "kind": "models"
      }
    ],
    "cost": {
      "zh": "本地生成占用自己的硬件、内存和磁盘；模型与工具的使用条件请看原始说明。",
      "en": "Local generation uses your own hardware, RAM and storage. See the original model and tool terms."
    },
    "applicableVersions": [
      "ComfyUI 0.30+ · native H3 templates"
    ],
    "difficulty": "beginner",
    "sourceRefs": [
      {
        "title": "Earlier community reference",
        "url": "https://x.com/ComfyUI/status/2084110113078751465"
      }
    ],
    "hardwareProfiles": [
      "vram-12",
      "vram-16",
      "vram-24-plus",
      "cloud-gpu"
    ],
    "checks": {
      "zh": [
        "使用 FL2VA，首尾帧连接与所选模式相符。",
        "成片可播放，输入图的关键构图得到保留。"
      ],
      "en": [
        "FL2VA is selected and boundary-frame connections match the intended mode.",
        "The result plays and retains the important input composition."
      ]
    },
    "expectedResult": {
      "zh": "让一张图片动起来，或用首尾两张图约束开场与结尾。",
      "en": "Animate a still image or constrain the opening and ending with two images."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "图片被裁切",
          "en": "Unexpected image crop"
        },
        "solution": {
          "zh": "先把输入图裁成目标宽高比，再对照工作流尺寸预览。",
          "en": "Crop the input to the target aspect ratio, then inspect the workflow preview dimensions."
        }
      },
      {
        "problem": {
          "zh": "把参考人物图当成固定首帧",
          "en": "A character reference became a literal first frame"
        },
        "solution": {
          "zh": "需要复用身份而不是固定构图时，改看 Ref2VA 教程。",
          "en": "Use the Ref2VA guide when you need identity reuse rather than a fixed opening composition."
        }
      }
    ],
    "nextGuideIds": [
      "video-pov-camera-angles",
      "prompt-agent-skill"
    ],
    "relatedCases": [
      {
        "id": "official-fl2va-ramen",
        "title": {
          "zh": "拉面与家庭晚餐：首帧控制示例",
          "en": "Ramen Rack Focus: a first-frame example"
        },
        "relationship": "example"
      }
    ],
    "learningOrder": 2,
    "guideType": "project",
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T12:45:10.151Z"
    }
  },
  {
    "id": "4080-int8-pruned",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "RTX 4080 16GB：Pruned + INT8 低显存部署",
      "en": "RTX 4080 16GB: pruned plus INT8 low-VRAM setup"
    },
    "outcome": {
      "zh": "在 16GB 显存上跑通 10 秒本地 H3 测试。",
      "en": "Run a ten-second local H3 test within 16GB VRAM."
    },
    "audience": {
      "zh": "4080 16GB 或相近显存、需要低显存路线的用户。",
      "en": "Users around 16GB VRAM who need a reduced-memory path."
    },
    "hardware": {
      "zh": "RTX 4080 16GB（原帖实测）；约 10 分钟生成 10 秒视频。",
      "en": "RTX 4080 16GB as reported by the source; roughly ten minutes for a ten-second clip."
    },
    "prerequisites": {
      "zh": [
        "ComfyUI 原生工作流",
        "匹配的 pruned 与 INT8 权重",
        "正确模型目录"
      ],
      "en": [
        "The native ComfyUI workflow",
        "Matching pruned and INT8 weights",
        "Correct model directories"
      ]
    },
    "steps": {
      "zh": [
        "记录原帖的 4080 16GB 结果只作为硬件快照，不把约 10 分钟耗时当成所有机器的承诺。",
        "从源码安装 DiffSynth-Studio，并额外安装 quant 依赖；使用独立 Python 环境，避免改坏现有 ComfyUI。",
        "打开低显存示例 MiniMax-H3-Int8-ConvRot-Pruned-FL2VA.py，先核对模型 ID、origin_file_pattern、vram_limit 与输出路径。",
        "先运行示例默认的短片和低画布；监控峰值显存、系统内存、下载与 VAE 解码，不要一开始改成 10 秒或原生高分辨率。",
        "确认 MP4 含视频与声音后，再固定 seed 逐步增加帧数或画布。一次只改一个规模参数，并保留 OOM 前最后成功配置。",
        "量化结果与原生基线使用相同 Prompt、seed 和输出规格做画质与声音对照，不能只以能跑通判断质量。"
      ],
      "en": [
        "Treat the source's RTX 4080 16GB result as a hardware snapshot, not a universal ten-minute promise.",
        "Install DiffSynth-Studio from source with its quant extras in a separate Python environment so the ComfyUI environment remains untouched.",
        "Open MiniMax-H3-Int8-ConvRot-Pruned-FL2VA.py under model_inference_low_vram and verify model IDs, origin_file_pattern, vram_limit, and output path.",
        "Run the example first at its short duration and small canvas. Monitor peak VRAM, system memory, downloads, and VAE decode before increasing duration or resolution.",
        "After an MP4 with video and audio succeeds, hold the seed constant and increase frames or canvas one variable at a time. Keep the last successful configuration before OOM.",
        "Compare quantized and native outputs with the same prompt, seed, and output specification; fitting in memory is not a quality verdict."
      ]
    },
    "commands": [
      "git clone https://github.com/modelscope/DiffSynth-Studio.git",
      "cd DiffSynth-Studio && pip install -e .",
      "pip install \"diffsynth[quant]\"",
      "python examples/minimax_h3/model_inference_low_vram/MiniMax-H3-Int8-ConvRot-Pruned-FL2VA.py"
    ],
    "checks": {
      "zh": [
        "量化依赖可以导入且低显存示例开始加载匹配的 FL2VA 权重",
        "一次短片成功输出含视频轨和声音轨的 MP4",
        "已记录峰值显存、总耗时、画布、帧数与 seed",
        "与原生基线的画面和声音差异已人工检查"
      ],
      "en": [
        "Quantization dependencies import and the low-VRAM example loads matching FL2VA weights",
        "A short run produces an MP4 with video and audio streams",
        "Peak VRAM, total time, canvas, frames, and seed are recorded",
        "Visual and audio differences from the native baseline are reviewed"
      ]
    },
    "caveats": {
      "zh": [
        "量化与剪枝可能影响质量；先复现，再做画质判断。"
      ],
      "en": [
        "Quantization and pruning may affect quality; reproduce first, then evaluate."
      ]
    },
    "posterUrl": "/tutorial-posters/4080-int8.jpg",
    "tags": [
      "RTX 4080",
      "INT8",
      "Pruned"
    ],
    "relatedResourceIds": [
      "awesome-h3-resources",
      "official-minimax-h3"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/eternityspring/status/2084666970558074971",
      "author": "Shuohao",
      "handle": "@eternityspring",
      "publishedAt": "2026-08-04",
      "originalLanguage": "zh"
    },
    "engagement": {
      "replies": 34,
      "reposts": 140,
      "likes": 734,
      "views": 180000,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": false,
    "difficulty": "intermediate",
    "estimatedMinutes": 120,
    "hardwareProfiles": [
      "vram-16",
      "cloud-gpu"
    ],
    "expectedResult": {
      "zh": "在 16GB 级显卡上完成一个短时长 INT8 pruned FL2VA 带声视频，并记录可复现的显存、耗时和质量基线。",
      "en": "Complete one short INT8 pruned FL2VA video with audio on a 16GB-class GPU and record reproducible VRAM, timing, and quality baselines."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "comfy-kitchen 或量化后端无法导入",
          "en": "comfy-kitchen or the quant backend will not import"
        },
        "solution": {
          "zh": "确认在 DiffSynth 独立环境安装 diffsynth[quant]，并让 Torch/CUDA 版本与显卡匹配。",
          "en": "Install diffsynth[quant] inside the isolated DiffSynth environment and match Torch/CUDA to the GPU."
        }
      },
      {
        "problem": {
          "zh": "加载成功但生成时 OOM",
          "en": "The model loads but generation OOMs"
        },
        "solution": {
          "zh": "回到低显存示例默认画布与帧数，减少 vram_limit 预留之外的并行程序，再逐级放大。",
          "en": "Return to the example's default canvas and frames, close competing GPU processes, and scale one step at a time."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "删除独立 DiffSynth 虚拟环境与仓库即可，不要在 ComfyUI 环境直接批量卸载依赖。"
      ],
      "en": [
        "Remove the isolated DiffSynth environment and repository; do not bulk-uninstall dependencies from the ComfyUI environment."
      ]
    },
    "sourceRefs": [
      {
        "title": "DiffSynth MiniMax H3 documentation",
        "url": "https://github.com/modelscope/DiffSynth-Studio/blob/main/docs/en/Model_Details/MiniMax-H3.md"
      }
    ],
    "learningTrack": "run",
    "depth": "guide",
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "applicableVersions": [
      "DiffSynth-Studio main @ 2026-08-23",
      "INT8 ConvRot pruned FL2VA low-VRAM example"
    ],
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "commandItems": [
      {
        "kind": "command",
        "value": "git clone https://github.com/modelscope/DiffSynth-Studio.git",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "cd DiffSynth-Studio && pip install -e .",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "pip install \"diffsynth[quant]\"",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "python examples/minimax_h3/model_inference_low_vram/MiniMax-H3-Int8-ConvRot-Pruned-FL2VA.py",
        "platform": "shell"
      }
    ],
    "original": {
      "kind": "x-article",
      "capturedAt": "2026-09-13T09:45:22.549Z"
    }
  },
  {
    "id": "four-bit-eight-gb",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "4-bit + DiffSynth：最低 8GB 显存路线",
      "en": "4-bit plus DiffSynth: an 8GB-VRAM route"
    },
    "outcome": {
      "zh": "通过 4-bit 量化与显存自适应推理在更小显存上运行 H3。",
      "en": "Run H3 on smaller GPUs through 4-bit quantization and VRAM-adaptive inference."
    },
    "audience": {
      "zh": "8GB 显存或希望在 Mac 上尝试量化 H3 的用户。",
      "en": "Users with 8GB VRAM or those testing a quantized H3 path on Mac."
    },
    "hardware": {
      "zh": "最低 8GB VRAM 为来源公布下限；实际速度和可用模式需复核。",
      "en": "The source reports an 8GB minimum; verify actual speed and supported modes."
    },
    "prerequisites": {
      "zh": [
        "DiffSynth-Studio",
        "匹配的 4-bit H3 权重",
        "足够系统内存与磁盘"
      ],
      "en": [
        "DiffSynth-Studio",
        "Matching 4-bit H3 weights",
        "Enough system memory and storage"
      ]
    },
    "steps": {
      "zh": [
        "新建独立 Python 环境，从源码安装 DiffSynth-Studio；先查看 H3 文档当前是否仍标注 NF4 最低约 7GB VRAM。",
        "打开 model_inference_low_vram/MiniMax-H3-NF4-FL2VA.py，保留四个 NF4 模型文件的匹配 origin_file_pattern。",
        "让 vram_limit 使用可用显存减 2GB 的文档默认策略；不要伪装或硬编码比物理显存更大的数值。",
        "第一轮使用 480×832、124 帧或示例更小设置，固定 seed，等待模型按需在 CPU/GPU 间调度。",
        "用 ffprobe 检查输出 t2va.mp4 的视频、24fps、32kHz 立体声音频轨，并人工试听同步。",
        "低显存基线通过后再分别尝试 Ref2VA、画布或帧数；记录每次总耗时和系统内存，避免只报告 GPU 显存。"
      ],
      "en": [
        "Create an isolated Python environment and install DiffSynth-Studio from source. Confirm the current H3 documentation still reports an approximately 7GB NF4 minimum.",
        "Open model_inference_low_vram/MiniMax-H3-NF4-FL2VA.py and preserve the four matching NF4 origin_file_pattern values.",
        "Keep the documented vram_limit strategy of available VRAM minus 2GB; never spoof a limit larger than physical memory.",
        "Start at 480×832 and 124 frames or the example's smaller setting, hold the seed constant, and allow CPU/GPU parameter movement to finish.",
        "Use ffprobe to confirm t2va.mp4 contains video, 24fps, and a 32kHz stereo audio stream, then listen for sync.",
        "After the low-VRAM baseline passes, test Ref2VA, canvas, or frame count separately. Record total time and system memory instead of reporting GPU VRAM alone."
      ]
    },
    "commands": [
      "git clone https://github.com/modelscope/DiffSynth-Studio.git",
      "cd DiffSynth-Studio && pip install -e .",
      "python examples/minimax_h3/model_inference_low_vram/MiniMax-H3-NF4-FL2VA.py",
      "ffprobe -v error -show_streams t2va.mp4"
    ],
    "checks": {
      "zh": [
        "NF4 四个组件与 FL2VA processor 成功加载",
        "任务没有伪造 vram_limit，且在物理显存范围内完成",
        "t2va.mp4 包含视频与 32kHz 立体声音频",
        "已记录 GPU 显存、系统内存、总耗时与量化质量差异"
      ],
      "en": [
        "All four NF4 components plus the FL2VA processor load",
        "The job completes within physical VRAM without spoofing vram_limit",
        "t2va.mp4 contains video and 32kHz stereo audio",
        "GPU VRAM, system memory, total time, and quantization differences are recorded"
      ]
    },
    "caveats": {
      "zh": [
        "低显存可运行不等于速度快；量化质量与原生模型不同。"
      ],
      "en": [
        "Running in low VRAM does not imply speed, and quantized quality differs from native weights."
      ]
    },
    "posterUrl": "/tutorial-posters/brand-low-vram.svg",
    "tags": [
      "4-bit",
      "DiffSynth",
      "8GB VRAM"
    ],
    "relatedResourceIds": [
      "awesome-h3-resources",
      "official-minimax-h3"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/ModelScope2022/status/2084625441940279770",
      "author": "ModelScope",
      "handle": "@ModelScope2022",
      "publishedAt": "2026-08-04",
      "originalLanguage": "en"
    },
    "engagement": {
      "replies": 33,
      "reposts": 169,
      "likes": 1498,
      "views": 132704,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": false,
    "difficulty": "intermediate",
    "estimatedMinutes": 150,
    "hardwareProfiles": [
      "vram-8",
      "vram-12"
    ],
    "expectedResult": {
      "zh": "在约 7–8GB 可用显存的 NVIDIA 环境完成一条 NF4 FL2VA 带声短片，并清楚记录速度与质量代价。",
      "en": "Complete one NF4 FL2VA short video with audio in an NVIDIA environment with roughly 7–8GB usable VRAM, with speed and quality costs documented."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "下载或载入 NF4 组件失败",
          "en": "An NF4 component fails to download or load"
        },
        "solution": {
          "zh": "逐一核对四个 origin_file_pattern 与模型 ID；FL2VA 和 Ref2VA 组件不能混用。",
          "en": "Check each of the four origin_file_pattern values and model IDs; do not mix FL2VA and Ref2VA components."
        }
      },
      {
        "problem": {
          "zh": "显存够但系统卡死",
          "en": "VRAM fits but the system stalls"
        },
        "solution": {
          "zh": "量化低显存依赖 CPU offload；检查系统内存和交换空间，减小画布或帧数后重试。",
          "en": "Low-VRAM quantization relies on CPU offload; inspect system RAM and swap, then reduce canvas or frames."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "删除独立 DiffSynth 环境与下载缓存前先确认其他模型未共用。"
      ],
      "en": [
        "Before removing the isolated DiffSynth environment and download cache, confirm no other model shares them."
      ]
    },
    "sourceRefs": [
      {
        "title": "DiffSynth MiniMax H3 documentation",
        "url": "https://github.com/modelscope/DiffSynth-Studio/blob/main/docs/en/Model_Details/MiniMax-H3.md"
      }
    ],
    "learningTrack": "run",
    "depth": "guide",
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "applicableVersions": [
      "DiffSynth-Studio main @ 2026-08-23",
      "MiniMax-H3-NF4 low-VRAM example"
    ],
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "commandItems": [
      {
        "kind": "command",
        "value": "git clone https://github.com/modelscope/DiffSynth-Studio.git",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "cd DiffSynth-Studio && pip install -e .",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "python examples/minimax_h3/model_inference_low_vram/MiniMax-H3-NF4-FL2VA.py",
        "platform": "shell"
      },
      {
        "kind": "command",
        "value": "ffprobe -v error -show_streams t2va.mp4",
        "platform": "shell"
      }
    ],
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:45:25.984Z"
    }
  },
  {
    "id": "train-ref2va-lora",
    "contentType": "community",
    "category": "training",
    "title": {
      "zh": "H3 Ref2VA LoRA：用 AI Toolkit 训练视频 LoRA",
      "en": "H3 Ref2VA LoRA: train a video LoRA with AI Toolkit"
    },
    "outcome": {
      "zh": "了解原作者的 Ref2VA LoRA 训练路线，先确认工具、数据格式与硬件要求。",
      "en": "Explore the author’s Ref2VA LoRA training route and identify its tool, dataset format and hardware requirements."
    },
    "audience": {
      "zh": "有训练经验、数据和高显存 GPU 的开发者。",
      "en": "Developers with training experience, data, and a high-memory GPU."
    },
    "hardware": {
      "zh": "训练硬件要求待原作者项目文档复查，不将其他训练项目的显存要求套用于本篇。",
      "en": "Training hardware requirements need review against the author’s project; requirements from another training project do not apply."
    },
    "prerequisites": {
      "zh": [
        "有权使用的训练数据",
        "先确认原帖所用工具与 H3 模型分支",
        "核对同一项目的安装与训练说明"
      ],
      "en": [
        "Training data you have permission to use",
        "Identify the tool and H3 model variant used in the source",
        "Check installation and training instructions from that same project"
      ]
    },
    "steps": {
      "zh": [
        "从原作者帖子确认所用训练工具和模型分支。",
        "到该工具文档核对数据格式、硬件与安装要求。",
        "只使用同一项目的训练与推理说明；不要混用其他项目的脚本。"
      ],
      "en": [
        "Identify the training tool and model variant in the original post.",
        "Check that tool documentation for data format, hardware and setup.",
        "Use training and inference instructions from the same project; do not mix scripts from another project."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "训练会消耗大量算力；数据版权和模型许可必须先确认。"
      ],
      "en": [
        "Training is compute-intensive; confirm dataset rights and model licensing first."
      ]
    },
    "posterUrl": "/tutorial-posters/train-ref2va.jpg",
    "tags": [
      "Training",
      "Ref2VA",
      "LoRA"
    ],
    "relatedResourceIds": [
      "official-minimax-h3"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/ostrisai/status/2090540475199680543",
      "author": "Ostris",
      "handle": "@ostrisai",
      "publishedAt": "2026-08-21",
      "originalLanguage": "en"
    },
    "engagement": {
      "replies": 14,
      "reposts": 33,
      "likes": 342,
      "views": 14000,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "flagship": false,
    "difficulty": "advanced",
    "estimatedMinutes": 240,
    "hardwareProfiles": [
      "vram-24-plus",
      "cloud-gpu"
    ],
    "learningTrack": "create",
    "depth": "guide",
    "evidence": {
      "status": "needs-review",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原帖介绍 AI Toolkit；旧版本站正文混入另一训练项目，现已移除。完整训练命令需回到对应项目核对。",
      "en": "The original post discusses AI Toolkit. The previous summary mixed in a different training project; those instructions have been removed. Check the matching project for training commands."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "nextGuideIds": [
      "video-pov-camera-angles"
    ],
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:45:38.788Z"
    }
  },
  {
    "id": "webui-motion-context",
    "contentType": "community",
    "category": "long-video",
    "title": {
      "zh": "H3 WebUI：Motion Context + 内置升频",
      "en": "H3 WebUI: Motion Context plus built-in upscaling"
    },
    "outcome": {
      "zh": "通过 Motion Context 把短片续接到最长约一分钟，并在导出前升频。",
      "en": "Chain short clips toward roughly one minute with Motion Context and upscale before export."
    },
    "audience": {
      "zh": "希望用 WebUI 而不是手接 ComfyUI 图的用户。",
      "en": "Users who prefer a WebUI over manually wiring a ComfyUI graph."
    },
    "hardware": {
      "zh": "Apple Silicon / 项目当前支持环境；以 WebUI 文档为准。",
      "en": "Apple Silicon or the environments currently supported by the WebUI."
    },
    "prerequisites": {
      "zh": [
        "H3 WebUI 1.4 或更新版本",
        "短片分段计划",
        "足够的生成和升频时间"
      ],
      "en": [
        "H3 WebUI 1.4 or newer",
        "A short-clip segment plan",
        "Enough time for generation and upscaling"
      ]
    },
    "steps": {
      "zh": [
        "先用 5 秒 720p 验证基础生成。",
        "开启 Motion Context 并分段生成，逐段检查接缝。",
        "最后再启用内置升频，不把升频问题混入生成调参。"
      ],
      "en": [
        "Validate a five-second 720p base generation.",
        "Enable Motion Context and generate in segments while checking each join.",
        "Apply built-in upscaling last so its artifacts do not confuse generation tuning."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "最长时长和加速比例来自原帖版本快照，升级后需重新核验。"
      ],
      "en": [
        "Maximum duration and speed claims are version snapshots; recheck after updates."
      ]
    },
    "posterUrl": "/tutorial-posters/brand-long-video.svg",
    "tags": [
      "WebUI",
      "Motion Context",
      "Upscale"
    ],
    "relatedResourceIds": [
      "motion-context",
      "antirez-h3-metal"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/core_tan/status/2087415961293144516",
      "author": "CALMDUST",
      "handle": "@core_tan",
      "publishedAt": "2026-08-12",
      "originalLanguage": "ja"
    },
    "engagement": {
      "replies": 6,
      "reposts": 28,
      "likes": 336,
      "views": 44501,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "create",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:46:51.496Z"
    }
  },
  {
    "id": "ref2v-motion-context-chunks",
    "contentType": "community",
    "category": "long-video",
    "title": {
      "zh": "4070 12GB：三段约 5 秒续接成 13 秒角色舞蹈",
      "en": "4070 12GB: chain three roughly 5-second chunks into a 13-second character dance"
    },
    "outcome": {
      "zh": "在单次 13 秒无法生成时，用 Turbo 8 步与 Motion Context 分块完成。",
      "en": "When a 13-second pass does not fit, use eight-step Turbo and Motion Context chunks."
    },
    "audience": {
      "zh": "12GB 显存、需要角色替换或长动作连续性的用户。",
      "en": "12GB users who need character replacement or longer motion continuity."
    },
    "hardware": {
      "zh": "RTX 4070 12GB + 32GB RAM（原帖实测）。",
      "en": "RTX 4070 12GB plus 32GB RAM, as reported by the source."
    },
    "prerequisites": {
      "zh": [
        "Ref2VA 工作流",
        "Turbo 8 步",
        "Motion Context",
        "动作参考视频"
      ],
      "en": [
        "A Ref2VA workflow",
        "Eight-step Turbo",
        "Motion Context",
        "A motion-reference video"
      ]
    },
    "steps": {
      "zh": [
        "把长动作切成三个约 5 秒区段。",
        "用 Ref2VA 保持角色与动作参考，使用 8 步 Turbo 生成。",
        "用 Motion Context 续接并检查动作、声音与身份边界。"
      ],
      "en": [
        "Split the longer motion into three roughly five-second sections.",
        "Use Ref2VA for character and motion reference with eight-step Turbo.",
        "Chain with Motion Context and inspect motion, audio, and identity at every join."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "原帖约 30 分钟是单一硬件实测；分块重叠和接缝需要人工检查。"
      ],
      "en": [
        "The roughly 30-minute time is hardware-specific; overlaps and joins require review."
      ]
    },
    "posterUrl": "/tutorial-posters/brand-motion.svg",
    "tags": [
      "RTX 4070",
      "Ref2VA",
      "Motion Context"
    ],
    "relatedResourceIds": [
      "motion-context",
      "turbo-lora"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/cla_kuro37459/status/2087892024410480796",
      "author": "Clakuro",
      "handle": "@cla_kuro37459",
      "publishedAt": "2026-08-13",
      "originalLanguage": "ja"
    },
    "engagement": {
      "replies": 3,
      "reposts": 145,
      "likes": 1256,
      "views": 101090,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "create",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "project",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:46:55.852Z"
    }
  },
  {
    "id": "lightx2v-vs-larryvrh",
    "contentType": "community",
    "category": "acceleration",
    "title": {
      "zh": "4 步与 8 步 LoRA：用 30 秒续接做 A/B",
      "en": "Four-step vs eight-step LoRA: a 30-second chaining A/B"
    },
    "outcome": {
      "zh": "比较 lightx2v 与 larryvrh LoRA 在清晰度、噪声、语音和稳定性上的差异。",
      "en": "Compare lightx2v and larryvrh LoRAs for sharpness, noise, speech, and stability."
    },
    "audience": {
      "zh": "已经能运行 Motion Context、正在选择 Turbo 权重的用户。",
      "en": "Users with Motion Context working who are choosing a Turbo weight."
    },
    "hardware": {
      "zh": "取决于 H3 WebUI 或 ComfyUI 当前路线。",
      "en": "Depends on the current H3 WebUI or ComfyUI path."
    },
    "prerequisites": {
      "zh": [
        "同一 Prompt、seed 与参考素材",
        "两套待比较 LoRA",
        "30 秒分段续接基线"
      ],
      "en": [
        "The same prompt, seed, and references",
        "Both LoRAs under test",
        "A 30-second chained baseline"
      ]
    },
    "steps": {
      "zh": [
        "固定全部输入和 Motion Context 参数。",
        "分别测试 4 步与 8 步，并记录 LoRA 强度。",
        "从清晰度、噪声、语音与接缝稳定性四项选择，不只看速度。"
      ],
      "en": [
        "Fix every input and Motion Context parameter.",
        "Test four and eight steps separately and record LoRA strength.",
        "Choose by sharpness, noise, speech, and join stability—not speed alone."
      ]
    },
    "commands": [],
    "caveats": {
      "zh": [
        "原作者结论属于其 WebUI 与当时权重版本，不能替代本机 A/B。"
      ],
      "en": [
        "The source's conclusion reflects its WebUI and weight versions; it does not replace a local A/B."
      ]
    },
    "posterUrl": "/tutorial-posters/brand-benchmark.svg",
    "tags": [
      "A/B Test",
      "4 Steps",
      "8 Steps"
    ],
    "relatedResourceIds": [
      "turbo-lora",
      "modeltc-h3-turbo",
      "motion-context"
    ],
    "source": {
      "platform": "x",
      "url": "https://x.com/core_tan/status/2088147718732734936",
      "author": "CALMDUST",
      "handle": "@core_tan",
      "publishedAt": "2026-08-14",
      "originalLanguage": "ja"
    },
    "engagement": {
      "replies": 1,
      "reposts": 4,
      "likes": 66,
      "views": 9745,
      "snapshotAt": "2026-08-23"
    },
    "verifiedAt": "2026-08-23",
    "addedAt": "2026-08-23T02:34:28+08:00",
    "learningTrack": "run",
    "depth": "guide",
    "flagship": false,
    "evidence": {
      "status": "active",
      "sourceCheckedAt": "2026-08-23"
    },
    "materialsNote": {
      "zh": "原作者资料见下方来源；未提供的素材或参数不由本站补写。",
      "en": "Follow the original source for its materials. Missing assets or settings are not reconstructed here."
    },
    "recommendation": {
      "zh": "先阅读适用条件，再到原作者内容中查看完整演示与更新。",
      "en": "Check the prerequisites, then follow the author for the full demonstration and updates."
    },
    "guideType": "reference",
    "original": {
      "kind": "x-thread",
      "capturedAt": "2026-09-13T09:47:14.883Z"
    }
  },
  {
    "id": "minimax-director-timeline",
    "chapters": [
      {
        "title": {
          "zh": "认识节点与输出检查",
          "en": "Inspect nodes and outputs"
        },
        "seconds": 91,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=91",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=91"
      },
      {
        "title": {
          "zh": "安排三条轨道",
          "en": "Arrange the three tracks"
        },
        "seconds": 282,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=282",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=282"
      },
      {
        "title": {
          "zh": "为角色配置对白",
          "en": "Assign character dialogue"
        },
        "seconds": 905,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=905",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=905"
      },
      {
        "title": {
          "zh": "约束首尾与中间画面",
          "en": "Constrain boundary and intermediate frames"
        },
        "seconds": 1106,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=1106",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=1106"
      },
      {
        "title": {
          "zh": "调节镜头运动",
          "en": "Adjust camera movement"
        },
        "seconds": 1312,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=1312",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=1312"
      },
      {
        "title": {
          "zh": "跨镜头复用主体",
          "en": "Reuse subjects across shots"
        },
        "seconds": 1428,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=1428",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=1428"
      },
      {
        "title": {
          "zh": "检查编译结果",
          "en": "Review compiled output"
        },
        "seconds": 1764,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=1764",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=1764"
      },
      {
        "title": {
          "zh": "跟随完整素材示例",
          "en": "Follow the complete media example"
        },
        "seconds": 1843,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=1843",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=1843"
      },
      {
        "title": {
          "zh": "查看生成结果",
          "en": "Inspect the rendered result"
        },
        "seconds": 1986,
        "url": "https://www.youtube.com/watch?v=XIKDfvoFcCY&t=1986",
        "urlZh": "https://www.youtube.com/watch?v=te8qpVaP0yI&t=1986"
      }
    ],
    "contentType": "community",
    "category": "comfyui",
    "title": {
      "zh": "MiniMax Director：用时间轴制作多镜头有声视频",
      "en": "MiniMax Director: multi-shot video with sound on a timeline"
    },
    "outcome": {
      "zh": "在 ComfyUI 中安排镜头、角色参考和对白，检查编译后的 Prompt，再输出有声 MP4。",
      "en": "Arrange shots, recurring subjects and dialogue in ComfyUI, inspect the compiled prompt, and render an MP4 with sound."
    },
    "audience": {
      "zh": "已经用 ComfyUI 官方 H3 模板生成过视频，希望学习多镜头与角色参考的创作者。",
      "en": "Creators who have already rendered an H3 clip with the official ComfyUI template and want to work with multiple shots and recurring subjects."
    },
    "hardware": {
      "zh": "生成需求由 H3 与所选工作流决定，插件本身主要编译提示词。作者在投稿中报告 Linux + NVIDIA H200 141GB 生成 8 秒约需 30 分钟；这不是最低配置，也不是本站实测。其他显卡耗时未知。",
      "en": "Rendering requirements come from H3 and the selected workflow; the plugin primarily compiles prompts. In the submission, the author reports roughly 30 minutes for 8 seconds on Linux with an NVIDIA H200 141GB. This is neither a minimum requirement nor a site benchmark; other GPUs were not timed."
    },
    "prerequisites": {
      "zh": [
        "先跑通 ComfyUI 官方 H3 模板；若尚未成功，先阅读本站 ComfyUI 入门路线。",
        "使用 ComfyUI 0.31.0 或更新版本，Python 3.10 或更新版本。",
        "准备 Comfy-Org/MiniMax-H3 的四个权重，保留原始文件名。作者 2026-09-02 统计约 39.6 GiB，另留输出与运行空间。"
      ],
      "en": [
        "Render one clip with the official ComfyUI H3 template first; use our ComfyUI setup guide if needed.",
        "Use ComfyUI 0.31.0 or newer and Python 3.10 or newer.",
        "Obtain the four Comfy-Org/MiniMax-H3 weights without renaming them. The author's September 2 snapshot totals about 39.6 GiB; allow additional space for outputs and runtime files."
      ]
    },
    "steps": {
      "zh": [
        "在 ComfyUI/custom_nodes 下执行下方克隆命令，然后重启 ComfyUI。",
        "将 minimax_h3_ref2va_pruned_int8_convrot.safetensors 放入 ComfyUI/models/diffusion_models/。",
        "将 qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors 放入 ComfyUI/models/text_encoders/；将 minimax_h3_video_vae_fp16.safetensors 和 minimax_h3_audio_vae_fp32.safetensors 放入 ComfyUI/models/vae/。",
        "打开 Workflow → Browse Templates → minimax-director；也可拖入插件的 example_workflows/minimax-director.json。先确认所有加载器都选中了正确文件，保持 Turbo 与 Upscale 关闭。",
        "选中播放头下的镜头块，填写该镜头的画面、运镜和声音；在时间轴安排后续镜头。时长会自动对齐 H3 允许的帧数。",
        "在 WHO & WHAT 为重复出现的人物或物体建立卡片，添加参考素材与描述；点击镜头 SUBJECTS 行的卡片，将 <Subject n> 插入镜头文本。",
        "检查 Prompt 面板的实际模型输入与 Report 面板的问题。修正明确报错后 Queue；查看输出 MP4 的画面、片长与声音。",
        "需要快速预览时，再准备作者文档中的 4-step Turbo LoRA，放入 models/loras/ 并开启 Speed 组的 Turbo。使用配套 4 步、强度 1.0 和 simple 调度；声音或快速动作变差时关闭 Turbo 回到普通生成。"
      ],
      "en": [
        "Run the clone command below from ComfyUI/custom_nodes, then restart ComfyUI.",
        "Place minimax_h3_ref2va_pruned_int8_convrot.safetensors in ComfyUI/models/diffusion_models/.",
        "Place qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors in ComfyUI/models/text_encoders/. Put minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors in ComfyUI/models/vae/.",
        "Open Workflow → Browse Templates → minimax-director, or drag in the plugin's example_workflows/minimax-director.json. Check every loader and leave Turbo and Upscale off initially.",
        "Select the shot block under the playhead and describe its picture, camera movement and sound. Arrange subsequent shots on the timeline; duration snaps to an accepted H3 frame count.",
        "Create WHO & WHAT cards for recurring people or objects with reference assets and descriptions. Click a card in the shot's SUBJECTS row to insert <Subject n> into the shot text.",
        "Inspect the actual model input in Prompt and resolve explicit errors in Report before queueing. Check the resulting MP4 for picture, duration and sound.",
        "For a faster preview, obtain the author's documented 4-step Turbo LoRA, place it in models/loras/ and enable Turbo in Speed. Use its matching four steps, strength 1.0 and simple scheduler. Disable Turbo if sound or fast motion deteriorates."
      ]
    },
    "commands": [
      "cd ComfyUI/custom_nodes\ngit clone https://github.com/imbutus/ComfyUI-MiniMaxDirector.git"
    ],
    "checks": {
      "zh": [
        "加载器能找到四个权重，Prompt 有内容，Report 没有阻断错误。",
        "生成的 MP4 能播放，片长符合时间轴，画面和声音均存在；自行核对对白同步与角色一致性。"
      ],
      "en": [
        "All four weights resolve, Prompt contains the compiled input, and Report has no blocking errors.",
        "The MP4 plays with the expected duration, picture and sound. Inspect dialogue synchronization and subject consistency yourself."
      ]
    },
    "caveats": {
      "zh": [
        "24 fps 下 H3 帧数满足 length % 17 == 5；编辑器会对齐时长，不等于可以无限续长。",
        "角色参考与对白编译不能保证每次生成都保持人物一致或精准口型。",
        "4-step Turbo 是预览方案，作者提示其音频与快速运动效果有局限。",
        "Upscale 默认关闭；开启前还需作者文档指定的上采样节点和权重，会增加计算量。"
      ],
      "en": [
        "At 24 fps, accepted H3 lengths satisfy length % 17 == 5. Timeline snapping does not enable unlimited duration.",
        "Reference cards and dialogue compilation cannot guarantee subject consistency or perfect lip sync in every generation.",
        "The four-step Turbo path is a preview option with author-reported audio and fast-motion limitations.",
        "Upscale is off by default. Enabling it requires the extra node pack and weights in the author's instructions and adds computation."
      ]
    },
    "posterUrl": "/tutorial-posters/minimax-director.webp",
    "tags": [
      "ComfyUI",
      "Ref2VA",
      "Timeline",
      "Multi-shot",
      "Dialogue"
    ],
    "relatedResourceIds": [],
    "source": {
      "platform": "github",
      "url": "https://github.com/imbutus/ComfyUI-MiniMaxDirector",
      "author": "imbutus",
      "originalLanguage": "en"
    },
    "verifiedAt": "2026-09-08",
    "learningTrack": "create",
    "depth": "deep",
    "difficulty": "intermediate",
    "recommendation": {
      "zh": "作者把安装、完整工作流、镜头编辑和排错整理成一套教程，适合从单镜头跨到多镜头；包含英文演示及中文配音入口。",
      "en": "The author provides installation, a complete graph, shot editing and troubleshooting in one tutorial, with an English walkthrough and a Chinese dub."
    },
    "cost": {
      "zh": "开源插件无需另购服务；实际视频生成仍消耗本地或云 GPU 算力，本站未购买算力复现实测。",
      "en": "The open-source plugin requires no separate service purchase. Rendering consumes local or rented GPU resources; we have not paid for a generation test."
    },
    "applicableVersions": [
      "MiniMax Director 0.17.14 (11fd464)",
      "ComfyUI >= 0.31.0",
      "Python >= 3.10"
    ],
    "learningResources": [
      {
        "label": {
          "zh": "完整工作流 JSON",
          "en": "Complete workflow JSON"
        },
        "url": "https://github.com/imbutus/ComfyUI-MiniMaxDirector/blob/main/example_workflows/minimax-director.json",
        "kind": "workflow"
      },
      {
        "label": {
          "zh": "四个 H3 权重文件",
          "en": "Four H3 weight files"
        },
        "url": "https://huggingface.co/Comfy-Org/MiniMax-H3",
        "kind": "models"
      },
      {
        "label": {
          "zh": "英文原版演示",
          "en": "English walkthrough"
        },
        "url": "https://youtu.be/XIKDfvoFcCY",
        "kind": "demo"
      },
      {
        "label": {
          "zh": "中文配音演示",
          "en": "Chinese dubbed walkthrough"
        },
        "url": "https://youtu.be/te8qpVaP0yI",
        "kind": "demo"
      },
      {
        "label": {
          "zh": "俄语配音演示",
          "en": "Russian dubbed walkthrough"
        },
        "url": "https://youtu.be/zjN37BQsOrI",
        "kind": "demo"
      },
      {
        "label": {
          "zh": "可选 Turbo LoRA",
          "en": "Optional Turbo LoRA"
        },
        "url": "https://huggingface.co/lightx2v/Minimax-h3-Turbo",
        "kind": "models"
      }
    ],
    "expectedResult": {
      "zh": "获得通过检查的多镜头 Prompt 和一段有声视频；实际画质、声音和角色连续性需观看结果判断。",
      "en": "A validated multi-shot prompt and a video with sound; inspect the output to judge image quality, audio and character continuity."
    },
    "troubleshooting": [
      {
        "problem": {
          "zh": "找不到示例工作流",
          "en": "Cannot find the example workflow"
        },
        "solution": {
          "zh": "当前路径为 example_workflows/minimax-director.json；旧投稿中的 examples/ 已过时，也可从 Browse Templates 打开。",
          "en": "Use example_workflows/minimax-director.json or Browse Templates. The examples/ path in the initial submission is outdated."
        }
      },
      {
        "problem": {
          "zh": "权重下拉框为空",
          "en": "Empty weight dropdown"
        },
        "solution": {
          "zh": "核对原始文件名和上述目录，刷新模型列表或重启；空下拉也可能是文件尚未下载完成。",
          "en": "Check original filenames and the directories above, then refresh or restart. Also confirm the download completed."
        }
      },
      {
        "problem": {
          "zh": "Turbo 画面正常但音频失真",
          "en": "Turbo picture looks fine but audio is distorted"
        },
        "solution": {
          "zh": "确认 ComfyUI >= 0.31.0 并提供 ModelSamplingAV；仍有问题先关闭 Turbo，作者也提示 4-step 音频有限制。",
          "en": "Confirm ComfyUI >= 0.31.0 with ModelSamplingAV. Disable Turbo if the problem persists; the author also notes four-step audio limitations."
        }
      }
    ],
    "uninstall": {
      "zh": [
        "先保存自己的工作流与素材。停止 ComfyUI，将 custom_nodes/ComfyUI-MiniMaxDirector 移出节点目录或移到废纸篓，再重启。",
        "H3 权重仍在 ComfyUI/models/，移除插件不会自动删除它们。"
      ],
      "en": [
        "Save your workflows and assets first. Stop ComfyUI, move custom_nodes/ComfyUI-MiniMaxDirector out of the node directory or to Trash, then restart.",
        "H3 weights remain under ComfyUI/models/; removing the plugin does not remove them."
      ]
    },
    "sourceRefs": [
      {
        "title": "MiniMax Director 0.17.14 README",
        "url": "https://github.com/imbutus/ComfyUI-MiniMaxDirector/blob/11fd464804d06730aafa7a419356b37a260fcac9/README.md"
      },
      {
        "title": "Original author submission #16",
        "url": "https://github.com/SkyNotSilent/awesome-MiniMax-H3-cases/issues/16"
      }
    ],
    "evidence": {
      "status": "active",
      "basis": "author-docs",
      "sourceCheckedAt": "2026-09-08"
    },
    "contribution": {
      "issueUrl": "https://github.com/SkyNotSilent/awesome-MiniMax-H3-cases/issues/16",
      "authorUrl": "https://github.com/imbutus"
    },
    "addedAt": "2026-09-08T12:39:39.434Z",
    "guideType": "project",
    "commandItems": [
      {
        "kind": "command",
        "value": "cd ComfyUI/custom_nodes\ngit clone https://github.com/imbutus/ComfyUI-MiniMaxDirector.git",
        "platform": "shell",
        "cwd": "Parent directory containing ComfyUI"
      }
    ],
    "original": {
      "kind": "markdown",
      "capturedAt": "2026-09-13T12:45:30.752Z"
    }
  }
]