"""Generate a batch of FastFrames-S clips: several seeds per prompt, umT5 loaded once. python batch.py prompts.txt out_dir [--seeds 0,1,2] [--reference] [--opts k=v,...] prompts file: one prompt per line (blank lines and lines starting with # are skipped), or JSON lines {"prompt": "...", "seeds": [1, 2], "name": "optional-slug"}; --seeds is the default list. Outputs out_dir/_s.mp4 + .json (timings), and out_dir/batch.json with all records. Same prompt + seed gives the same clip as run.py. MP4 encoding runs on a CPU thread while the GPU starts the next clip. Existing outputs are skipped (pass --overwrite to redo them). Weights are found as in run.py (--weights, $FASTFRAMES_MLX_WEIGHTS, this directory, or the Hub). """ import argparse import json import re import threading import time from pathlib import Path import mlx.core as mx import run def slug(text, n=48): s = re.sub(r"[^a-z0-9]+", "-", text.lower()).strip("-") return (s[:n].rstrip("-") or "prompt") def read_prompts(path, default_seeds): jobs = [] for line in Path(path).read_text().splitlines(): line = line.strip() if not line or line.startswith("#"): continue if line.startswith("{"): d = json.loads(line) jobs.append((d["prompt"], d.get("seeds", default_seeds), d.get("name") or slug(d["prompt"]))) else: jobs.append((line, default_seeds, slug(line))) names = [j[2] for j in jobs] for i, (p, s, n) in enumerate(jobs): # two prompts with the same slug: suffix if names.count(n) > 1: jobs[i] = (p, s, f"{n}-{run.emb_key(p)[:6]}") return jobs def main(): ap = argparse.ArgumentParser() ap.add_argument("prompts") ap.add_argument("out_dir") ap.add_argument("--seeds", default="0,1,2") ap.add_argument("--weights", default=None, help="directory with the .safetensors files") ap.add_argument("--frames", type=int, default=21) ap.add_argument("--decoder-dtype", default="fp32", choices=["fp32", "bf16"]) ap.add_argument("--reference", action="store_true") ap.add_argument("--opts", default="") ap.add_argument("--overwrite", action="store_true") a = ap.parse_args() out = Path(a.out_dir) out.mkdir(parents=True, exist_ok=True) wd = run.resolve_weights(a.weights) jobs = read_prompts(a.prompts, [int(s) for s in a.seeds.split(",")]) todo = [(p, seed, out / f"{name}_s{seed}.mp4") for p, seeds, name in jobs for seed in seeds] if not a.overwrite: todo = [j for j in todo if not j[2].exists()] print(f"{len(todo)} clips to make ({len(jobs)} prompts)", flush=True) if not todo: return t0 = time.perf_counter() embs, info = run.encode_prompts([p for p, _, _ in todo], wd, a.reference) text_s = round(time.perf_counter() - t0, 2) print(f"text: {text_s}s {info}", flush=True) t0 = time.perf_counter() ff = run.load_models(wd, a.reference, a.frames, a.decoder_dtype, run.parse_opts(a.opts)) load_s = round(time.perf_counter() - t0, 2) records, writer = [], None t_all = time.perf_counter() for i, (prompt, seed, path) in enumerate(todo): frames, rec = run.generate(ff, embs[prompt], seed, a.frames) rec = {"prompt": prompt, "seed": seed, "file": path.name, "path": "reference" if a.reference else "fast", **rec} if writer is not None: writer.join() def write(frames=frames, path=path, rec=rec): t = time.perf_counter() run.write_mp4(frames, path) rec["mp4_s"] = round(time.perf_counter() - t, 2) path.with_suffix(".json").write_text(json.dumps(rec, indent=1)) writer = threading.Thread(target=write) writer.start() records.append(rec) print(f"[{i + 1}/{len(todo)}] {path.name}: rollout {rec['rollout_s']}s decode {rec['decode_s']}s", flush=True) del frames mx.clear_cache() writer.join() summary = {"clips": len(records), "text_s": text_s, "load_s": load_s, **info, "gen_wall_s": round(time.perf_counter() - t_all, 2), "records": records} summary["per_clip_s"] = round(summary["gen_wall_s"] / len(records), 2) (out / "batch.json").write_text(json.dumps(summary, indent=1)) print(json.dumps({k: v for k, v in summary.items() if k != "records"})) if __name__ == "__main__": main()