#!/usr/bin/env python3 # -*- coding: utf-8 -*- """h3_official_shot.py — 用**官方 MinimaxH3Director 工作流**（无任何加速节点/LoRA）逐镜生成 MV 片段。 严格遵循的三条运行纪律（用户 2026-09-25 明确要求）： 1. 只用官方工作流（MiniMaxH3Director + ref2va/fl2va 官方权重 + 25 步 res_multistep/simple）， **绝不使用** Director-加速版 / 二采加速 / DSW V18 等含 MiniMaxH3MemoryEfficientSageAttentionPatch 或 turbo LoRA 的加速链路（那是 GPU 掉总线的元凶）。 2. 单镜时长 ≤ 11 秒（帧数取 17k+5 网格）。 3. 每个生成任务结束后强制冷却 90 秒，让 GPU 释放显存与降温。 用法: python3 h3_official_shot.py --shot S02 [--size 960x544] [--dur 5.2] [--seed 123] \ [--prompt-file <文件> | --keyframe <图>] [--extra-ref <图>] [--dry-run] [--cooldown 90] """ import argparse, json, os, shutil, subprocess, sys, time from pathlib import Path sys.path.insert(0, "/home/zyw/Downloads/dasiwa-h3-research") import v18_pipeline as p # 仅复用上传/提交/轮询/下载这几个工具函数（不碰它的加速模板） ROOT = Path("/home/zyw/Downloads/dl-hub/54-九九八十一MV") OUTBASE = ROOT / "06-H3片段" OUTBASE.mkdir(parents=True, exist_ok=True) TASK_R2V = "r2v — 参考主体生视频(Reference to Video)" TASK_T2V = "t2v — 文生视频(Text to Video)" UNET_R2V = "minimax_h3_ref2va_pruned_int8_convrot.safetensors" UNET_TEXT = "minimax_h3_fl2va_pruned_int8_convrot.safetensors" CLIP = "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors" VAE_VIDEO = "minimax_h3_video_vae_fp16.safetensors" VAE_AUDIO = "minimax_h3_audio_vae_fp32.safetensors" STEPS = 12 # 用户指定：12 步足够（官方流 25 步为基准，实测步数与耗时近线性） SAMPLER = "res_multistep" SCHEDULER = "simple" CFG = 1.0 SHIFT_VIDEO = 12.0 SHIFT_AUDIO = 3.0 FPS = 24.0 MAX_SEC = 11.0 MIN_FRAMES = 124 # 17*7+5 = 5.17s，H3 官方流实测可用的最短稳定档 def frames_for(dur): """取 ≤ dur 秒的最大 17k+5 帧数（≤11s）。""" want = min(dur, MAX_SEC) * FPS k = int((want - 5) // 17) n = 17 * k + 5 while n > MAX_SEC * FPS: k -= 1 n = 17 * k + 5 return max(n, MIN_FRAMES) def build_timeline(task_label, prompt_text, remote_refs, w, h, frames): refs = [{"index": i, "imageFile": fn, "subfolder": "", "type": "input"} for i, fn in enumerate(remote_refs or [])] return { "version": 4, "editMode": "global", "timelineMode": "gen_blank", "totalFrames": frames, "frameRate": FPS, "width": w, "height": h, "refMaxSize": max(w, h), "output": {"mode": "fixed", "longEdge": max(w, h), "width": w, "height": h, "maxExportFrames": 0, "exportMode": "all", "continuityEnabled": False, "continuityOverlapFrames": 9}, "videoClips": [], "video": {"fileName": "", "videoFile": "", "subfolder": "", "type": "input", "frames": [], "frameMap": []}, "global": {"taskType": task_label, "prompt": prompt_text, "commonEnabled": True, "refs": refs, "referenceVideo": {}, "continuousReference": False, "genImage": {"imageFile": ""}}, "segments": [{"id": "s0", "start": 0, "length": frames, "frameCount": frames, "durationSec": round(frames / FPS, 3), "prompt": "", "taskType": "", "refs": [], "referenceVideo": {}, "genImage": {"imageFile": ""}, "negativePrompt": ""}], "gen": {"defaultFrameCount": frames}, "runSelectEnabled": False, "runSelection": [], } def build_api(prompt_text, remote_refs, seed, task, w, h, frames): unet = UNET_R2V if task == "r2v" else UNET_TEXT label = TASK_R2V if task == "r2v" else TASK_T2V timeline = build_timeline(label, prompt_text, remote_refs, w, h, frames) return { "1": {"class_type": "UNETLoader", "inputs": {"unet_name": unet, "weight_dtype": "default"}}, "2": {"class_type": "CLIPLoader", "inputs": {"clip_name": CLIP, "type": "minimax", "device": "default"}}, "3": {"class_type": "VAELoader", "inputs": {"vae_name": VAE_VIDEO}}, "4": {"class_type": "VAELoader", "inputs": {"vae_name": VAE_AUDIO}}, "5": {"class_type": "MiniMaxH3Director", "inputs": { "model": ["1", 0], "video_vae": ["3", 0], "audio_vae": ["4", 0], "clip": ["2", 0], "task_type": label, "global_prompt": prompt_text, "bd_grp_sample": "采样设置", "cfg": CFG, "seed": seed, "frame_rate": FPS, "width": w, "height": h, "ref_max_size": max(w, h), "total_frames": frames, "timeline_data": json.dumps(timeline, ensure_ascii=False), "bd_grp_advanced": "高级采样", "steps": STEPS, "sampler": SAMPLER, "scheduler": SCHEDULER, "shift_video": SHIFT_VIDEO, "shift_audio": SHIFT_AUDIO, "bd_grp_perf": "性能 Performance", "clear_vram_between_segments": True, "export_source_images": False}}, "6": {"class_type": "CreateVideo", "inputs": {"images": ["5", 0], "fps": ["5", 2], "audio": ["5", 1]}}, "7": {"class_type": "SaveVideo", "inputs": {"video": ["6", 0], "filename_prefix": "video/OfficialDirector", "format": "auto", "codec": "h264"}}, } def gpu_temp(): """读 GPU 机温度（通过 ssh_exec.sh，失败返回 None）。""" try: r = subprocess.run(["bash", os.path.expanduser("~/Downloads/ssh_exec.sh"), "nvidia-smi --query-gpu=temperature.gpu --format=csv,noheader"], capture_output=True, text=True, timeout=60) for line in r.stdout.splitlines(): line = line.strip() if line.isdigit(): return int(line) except Exception: pass return None def main(): ap = argparse.ArgumentParser() ap.add_argument("--shot", required=True) ap.add_argument("--prompt-file", default=None) ap.add_argument("--prompt", default=None) ap.add_argument("--keyframe", default=None, help="主参考图（通常是该镜关键帧）") ap.add_argument("--extra-ref", action="append", default=[], help="附加参考图（如角色设定图）") ap.add_argument("--size", default="1344x768") ap.add_argument("--dur", type=float, default=5.2) ap.add_argument("--seed", type=int, default=20260925) ap.add_argument("--task", default="auto", choices=["auto", "r2v", "t2v"]) ap.add_argument("--cooldown", type=int, default=300, help="段间冷却秒数（记录①：13 小时零挂机配方 = 段间 5 分钟）") ap.add_argument("--max-temp", type=int, default=70, help="冷却后仍需等待 GPU 降到该温度再开下一镜") ap.add_argument("--worker", default="kf", help="输出子目录（默认关键帧目录名）") ap.add_argument("--dry-run", action="store_true") a = ap.parse_args() w, h = (int(x) for x in a.size.lower().split("x")) frames = frames_for(a.dur) prompt = (Path(a.prompt_file).read_text(encoding="utf-8").strip() if a.prompt_file else (a.prompt or "")) if not prompt: sys.exit("需要 --prompt-file 或 --prompt") task = a.task if task == "auto": task = "r2v" if a.keyframe else "t2v" shot_dir = OUTBASE / a.shot shot_dir.mkdir(parents=True, exist_ok=True) (shot_dir / "prompt.txt").write_text(prompt, encoding="utf-8") refs_local = ([Path(a.keyframe)] if a.keyframe else []) + [Path(x) for x in a.extra_ref] remote = [] for i, lp in enumerate(refs_local, 1): rn = f"{a.shot}_ref{i}_{lp.name}" if not a.dry_run: p.upload_image_to_comfy(str(lp), rn) remote.append(rn) print(f"[{a.shot}] 参考图 {remote} | {w}x{h} | {frames}帧={frames/FPS:.2f}s | steps={STEPS} | seed={a.seed} | task={task}", flush=True) api = build_api(prompt, remote, a.seed, task, w, h, frames) (shot_dir / "api_prompt.json").write_text(json.dumps(api, ensure_ascii=False, indent=1), encoding="utf-8") if a.dry_run: print("dry-run：已写出 api_prompt.json"); return 0 t0 = time.time() entry = p.submit_and_wait(api, timeout_min=180) info = p.extract_video_output(entry) video = p.download_video(info, shot_dir) final = shot_dir / f"{a.shot}.mp4" if final.exists(): final.unlink() shutil.move(str(video), str(final)) dt = time.time() - t0 print(f"[{a.shot}] 完成 {final} | 生成耗时 {dt/60:.1f} 分钟", flush=True) with open(ROOT / "logs" / "h3_official.log", "a", encoding="utf-8") as f: f.write(json.dumps({"shot": a.shot, "size": a.size, "frames": frames, "steps": STEPS, "seed": a.seed, "task": task, "seconds": round(dt, 1), "output": str(final)}, ensure_ascii=False) + "\n") if a.cooldown > 0: print(f"[{a.shot}] 冷却 {a.cooldown}s（释放显存/降温）…", flush=True) time.sleep(a.cooldown) # 温度闸门：等 GPU 降到 max_temp 以下再允许下一镜（记录显示 80℃ 长期满载会导致硬挂） for k in range(10): t = gpu_temp() if t is None or t <= a.max_temp: print(f"[{a.shot}] 温度闸门通过：{t}℃ ≤ {a.max_temp}℃", flush=True) break print(f"[{a.shot}] 当前 {t}℃ > {a.max_temp}℃，继续等 30s…", flush=True) time.sleep(30) return 0 if __name__ == "__main__": sys.exit(main())