🎵 Stable Audio 3 Medium — KAP 集成验证

RTX 3090 · ComfyUI v0.30.1 · LCM 8-step · 2026-08-05

~10s

生成耗时 (5s 音频)

6.5GB

峰值显存

44.1kHz

立体声

8步

LCM 采样

SFX seed 42 5s

cinematic whoosh, air swoosh, transition sound effect, clean and punchy

SFX seed 100 5s

heavy metal door slam, deep impact, reverb tail, industrial

Ambient seed 7 10s

gentle rain on a tin roof, continuous, ambient, peaceful

One-shot seed 33 5s

electric guitar power chord, distorted, sustained, rock, single hit

📋 API 调用示例

POST /api/v1/stableaudio/generate
Content-Type: application/json

{
  "prompt": "cinematic whoosh, air swoosh...",
  "seconds_total": 5,
  "seed": 42,
  "steps": 8,
  "sampler_name": "lcm",
  "scheduler": "simple",
  "cfg": 1,
  "format": "mp3"
}

🔧 技术细节

模型: stable_audio_3_medium.safetensors (8.6GB, 1.4B params)
文本编码器: t5gemma_b_b_ul2.safetensors (1.1GB)
采样器: LCM 8-step + simple scheduler + cfg=1 (蒸馏模型)
秒数修复: ConditioningStableAudio 节点 (PR #14858)
最大时长: 380s (~6:20)
新增路由: /api/v1/stableaudio/generate + /api/v1/stableaudio/models