#!/usr/bin/env python3 """H3 ref2va direct ComfyUI API - REF2VA model + correct CLIP (qwen3vl_32b), no Turbo LoRA.""" import requests, json, os, time, sys COMFYUI = "http://localhost:8188" KAP = "http://localhost:10588" # ── Workflow builder ── def build_ref2va_workflow(prompt, width, height, length, ref_image_files): """ Build a ref2va workflow using: - REF2VA model: minimax_h3_ref2va_pruned_int8_convrot.safetensors - CLIP: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (NOT gemma!) - No Turbo LoRA (incompatible with REF2VA) - 20 steps standard """ wf = { "10": { "class_type": "UNETLoader", "inputs": { "unet_name": "minimax_h3_ref2va_pruned_int8_convrot.safetensors", "weight_dtype": "default" } }, "11": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors", "type": "minimax" } }, "12": { "class_type": "VAELoader", "inputs": { "vae_name": "minimax_h3_video_vae_fp16.safetensors" } }, "13": { "class_type": "VAELoader", "inputs": { "vae_name": "minimax_h3_audio_vae_fp32.safetensors" } }, # ref_image LoadImage nodes: 14, 140, 141, ... "20": { "class_type": "MiniMaxH3ReferenceToVideo", "inputs": { "clip": ["11", 0], "vae": ["12", 0], "audio_vae": ["13", 0], "prompt": prompt, "width": width, "height": height, "length": length, "ref_image_size": "match", # ref_image_N connections added below } }, "30": { "class_type": "MiniMaxH3SigmaShift", "inputs": { "model": ["10", 0], "shift_video": 12.0, "shift_audio": 3.0 } }, "90": { "class_type": "MiniMaxH3ReferenceToVideo", "inputs": { "clip": ["11", 0], "vae": ["12", 0], "audio_vae": ["13", 0], "prompt": "blurry, low quality, distorted, watermark", "width": width, "height": height, "length": length, "ref_image_size": "match" } }, "50": { "class_type": "KSampler", "inputs": { "model": ["30", 0], "positive": ["20", 0], "negative": ["90", 0], "latent_image": ["20", 1], "seed": 42, "steps": 20, "cfg": 7.0, "sampler_name": "euler", "scheduler": "normal", "denoise": 1.0 } }, "100": { "class_type": "VAEDecode", "inputs": { "samples": ["50", 0], "vae": ["12", 0] } }, "110": { "class_type": "SaveAnimatedWEBP", "inputs": { "images": ["100", 0], "filename_prefix": "h3_ref2va", "fps": 24.0, "lossless": False, "quality": 90, "method": "default" } } } # Add ref_image LoadImage nodes and connect to node 20 for i, fname in enumerate(ref_image_files): node_id = str(14 + i * 10) # 14, 24, 34, ... wf[node_id] = { "class_type": "LoadImage", "inputs": { "image": os.path.basename(fname) } } wf["20"]["inputs"][f"ref_image_{i+1}"] = [node_id, 0] return wf def upload_image(filepath): """Upload image to ComfyUI input directory.""" fname = os.path.basename(filepath) with open(filepath, "rb") as f: resp = requests.post(f"{COMFYUI}/upload/image", files={"image": f}, data={"overwrite": "true"}) if resp.status_code != 200: print(f" Upload failed: {resp.text}") return None return resp.json().get("name", fname) def submit_and_wait(wf, timeout=600): """Submit workflow and poll for completion.""" resp = requests.post(f"{COMFYUI}/prompt", json={"prompt": wf}) r = resp.json() if resp.status_code != 200 or "node_errors" in r: print(f" Validation error: {json.dumps(r, indent=2)[:500]}") return None pid = r.get("prompt_id") if not pid: print(f" No prompt_id: {r}") return None print(f" promptId: {pid}") start = time.time() while time.time() - start < timeout: time.sleep(5) s = requests.get(f"{COMFYUI}/history/{pid}").json() if pid in s: outputs = s[pid].get("outputs", {}) status = s[pid].get("status", {}) if status.get("completed"): for nid, out in outputs.items(): if "images" in out: return out["images"][0] if "gifs" in out: return out["gifs"][0] return None elif status.get("status_str") == "error": msgs = status.get("messages", []) print(f" ERROR: {msgs[:200] if msgs else 'unknown'}") return None elapsed = int(time.time() - start) print(f" ...running ({elapsed}s)") print(f" TIMEOUT after {timeout}s") return None def download_output(item, output_dir): """Download output from ComfyUI.""" fname = item["filename"] subfolder = item.get("subfolder", "") url = f"{COMFYUI}/view?filename={fname}&subfolder={subfolder}&type=output" resp = requests.get(url) if resp.status_code == 200: outpath = os.path.join(output_dir, fname) with open(outpath, "wb") as f: f.write(resp.content) return outpath return None def main(): base_dir = "/home/kai/文档/LTX/validition_v1" output_dir = "/home/kai/shared/h3_ref2va_results" os.makedirs(output_dir, exist_ok=True) cases = sorted([d for d in os.listdir(base_dir) if os.path.isdir(os.path.join(base_dir, d))]) for case_id in cases: case_dir = os.path.join(base_dir, case_id) print(f"\n{'='*60}") print(f"Case {case_id}: ref2va (REF2VA model + qwen3vl CLIP)") # Load prompt prompt_file = os.path.join(case_dir, "prompt.txt") if not os.path.exists(prompt_file): prompt_file = os.path.join(case_dir, "prompt_nsfw.txt") if not os.path.exists(prompt_file): print(f" No prompt file, skipping") continue with open(prompt_file) as f: prompt = f.read().strip() # Find images imgs = sorted([f for f in os.listdir(case_dir) if f.endswith((".png", ".jpg", ".webp"))]) if not imgs: print(f" No images, skipping") continue # Upload all images print(f" Images: {imgs}") uploaded = [] for img_name in imgs: path = os.path.join(case_dir, img_name) uploaded_name = upload_image(path) if uploaded_name: uploaded.append((uploaded_name, path)) if not uploaded: print(f" Upload failed, skipping") continue # Determine dimensions from workflow json if available wf_file = os.path.join(case_dir, "ltx_msr_workflow.json") width, height = 1344, 768 # default landscape if os.path.exists(wf_file): try: with open(wf_file) as f: wf_data = json.load(f) for node in wf_data.values(): if node.get("class_type") == "EmptyLatentVideo" or "width" in node.get("inputs", {}): w = node["inputs"].get("width") h = node["inputs"].get("height") if w and h: width, height = w, h break except: pass # Align to n%17==5 length = 124 # ~5s default while length % 17 != 5: length += 1 print(f" Prompt: {len(prompt)} chars") print(f" Dimensions: {width}x{height}, length={length}") print(f" Refs: {len(uploaded)}") # Build workflow with uploaded filenames wf = build_ref2va_workflow(prompt, width, height, length, [u[0] for u in uploaded]) # Submit print(f" Submitting...") result = submit_and_wait(wf, timeout=600) if result: outpath = download_output(result, output_dir) if outpath: print(f" ✅ DONE: {os.path.basename(outpath)}") else: print(f" ❌ Download failed") else: print(f" ❌ Generation failed") print(f"\n{'='*60}") print("ALL_DONE") if __name__ == "__main__": main()