#!/usr/bin/env python3 """Generate Qwen 2509/Nunchaku FP4 T2I references for the P7-5.2 frontal anchor. This is intentionally separate from the P7-5.3 full-body reference pilot: it has no image input and must not silently become an image-edit experiment. The ``edit`` filename token identifies the P7-5 editing workflow, not an image-to-image operation in this generator. The generated PNG and result JSON record the exact input contracts and runtime. """ from __future__ import annotations import argparse import hashlib import importlib.metadata import json import platform import time from pathlib import Path import torch from diffusers import QwenImagePipeline from huggingface_hub import snapshot_download from nunchaku import NunchakuQwenImageTransformer2DModel ASSETS = Path(__file__).resolve().parent IDENTITY_CONTRACT = ASSETS / "p7-5-2-character-identity-contract.json" STYLE_CONTRACT = ASSETS / "p7-5-2-face-style-prompt-contract.json" ILLUSTRATION_CONTRACT = ASSETS / "p7-5-2-face-illustration-prompt-contract.json" MODEL_ID = "Qwen/Qwen-Image" TRANSFORMER_REPOSITORY = "nunchaku-tech/nunchaku-qwen-image" TRANSFORMER_FILENAME = "svdq-fp4_r128-qwen-image.safetensors" TRANSFORMER_ID = f"{TRANSFORMER_REPOSITORY}/{TRANSFORMER_FILENAME}" HF_HUB_CACHE = ASSETS.parents[3] / ".tmp" / "download" / "huggingface" / "hub" # Output filenames identify their generator and run; do not create a separate # output directory. OUTPUT_DIR = ASSETS DEFAULT_STEPS = 10 SIZE = (1024, 1024) FRAMING_PROMPTS = { "head": ( "Strict frontal head-and-neck studio reference of one young East Asian woman in her early twenties, complete hair crown visible with clear " "empty margin above it and no hair cropped by any image edge, face centered and facing the camera " "with both eyes and ears visible: compact oval face, a defined high nose bridge, and a refined straight nose line; high-volume petrol-teal " "jaw-length bob with asymmetric fringe, loose S-waves, inward-curled ends, and side locks wider than the neck. Long slender gently upturned " "eyes; equal orange-amber irises with distinct centered round dark pupils, separate from eyelids and eyeliner. Show ears and neck. " "No text, accessory, panel, collage, or background scene." ), "torso": ( "Strict frontal head-and-upper-torso studio reference of one young East Asian woman in her early twenties, framed from the complete hair crown " "through the mid-chest with clear empty margin above the hair and no head, shoulder, or chest cropped by any image edge. Face centered and facing " "the camera with both eyes and ears visible: compact oval face, a defined high nose bridge, and a refined straight nose line; high-volume petrol-teal " "jaw-length bob with asymmetric fringe, loose S-waves, inward-curled ends, and side locks wider than the neck. Long slender gently upturned eyes; " "equal orange-amber irises with distinct centered round dark pupils, separate from eyelids and eyeliner. Show both shoulders, neck, collarbones, and " "the upper torso wearing a simple plain dark crew-neck top. No text, accessory, panel, collage, or background scene." ), } def sha256(path: Path) -> str: digest = hashlib.sha256() with path.open("rb") as source: for block in iter(lambda: source.read(1024 * 1024), b""): digest.update(block) return digest.hexdigest() def asset_record(path: Path) -> dict[str, str]: return {"path": str(path), "sha256": sha256(path)} def runtime_record() -> dict[str, object]: packages: dict[str, str] = {} for name in ("nunchaku", "diffusers", "torch", "transformers", "accelerate"): try: packages[name] = importlib.metadata.version(name) except importlib.metadata.PackageNotFoundError: packages[name] = "not-installed" return { "python": platform.python_version(), "platform": platform.platform(), "packages": packages, "cuda_available": torch.cuda.is_available(), "cuda_device": torch.cuda.get_device_name(0) if torch.cuda.is_available() else None, } def load_pipeline() -> QwenImagePipeline: transformer_path = Path( snapshot_download(TRANSFORMER_REPOSITORY, cache_dir=HF_HUB_CACHE, local_files_only=True) ) / TRANSFORMER_FILENAME model_path = Path(snapshot_download(MODEL_ID, cache_dir=HF_HUB_CACHE, local_files_only=True)) transformer = NunchakuQwenImageTransformer2DModel.from_pretrained(transformer_path) transformer.set_offload(True, use_pin_memory=True, num_blocks_on_gpu=1) pipe = QwenImagePipeline.from_pretrained( model_path, transformer=transformer, torch_dtype=torch.bfloat16, local_files_only=True, ) pipe._exclude_from_cpu_offload.append("transformer") pipe.enable_sequential_cpu_offload() return pipe def main() -> None: parser = argparse.ArgumentParser(description=__doc__) parser.add_argument( "--framing", choices=tuple(FRAMING_PROMPTS), default="head", help="Use 'torso' to include both shoulders and the upper torso through the mid-chest.", ) parser.add_argument("--seed", type=int, default=62294) parser.add_argument("--steps", type=int, default=DEFAULT_STEPS) parser.add_argument( "--size", type=int, default=SIZE[0], help="Square output size in pixels; use a smaller value only when GPU memory prevents generation.", ) parser.add_argument("--run-label", default="front-1024-reference-v1") parser.add_argument("--output-dir", type=Path, default=OUTPUT_DIR) args = parser.parse_args() if args.steps < 1: raise ValueError("--steps must be at least 1") if args.size < 256 or args.size % 16: raise ValueError("--size must be at least 256 and divisible by 16") if not torch.cuda.is_available(): raise RuntimeError("CUDA is required") if missing := [path for path in (IDENTITY_CONTRACT, STYLE_CONTRACT, ILLUSTRATION_CONTRACT) if not path.is_file()]: raise FileNotFoundError("missing P7-5.2 face contract: " + ", ".join(map(str, missing))) illustration_contract = json.loads(ILLUSTRATION_CONTRACT.read_text(encoding="utf-8")) illustration_prompt = illustration_contract["front_face_illustration_prompt"] prompt = f"{illustration_prompt} {FRAMING_PROMPTS[args.framing]}" output_dir = args.output_dir if args.output_dir.is_absolute() else ASSETS / args.output_dir output_dir.mkdir(parents=True, exist_ok=True) stem = f"p7-5-2-qwen-face-{args.framing}-{args.run_label}-seed-{args.seed}-steps-{args.steps}-size-{args.size}" output = output_dir / f"{stem}.png" result_record = output_dir / f"{stem}-result.json" started = time.monotonic() pipeline = load_pipeline() image = pipeline( prompt=prompt, generator=torch.Generator("cpu").manual_seed(args.seed), true_cfg_scale=4.0, guidance_scale=1.0, negative_prompt=" ", num_inference_steps=args.steps, width=args.size, height=args.size, ).images[0] image.save(output) record = { "status": "generated", "experiment_id": f"p7-5-2-qwen-face-{args.framing}", "model": MODEL_ID, "transformer": TRANSFORMER_ID, "runtime": runtime_record(), "identity_contract": asset_record(IDENTITY_CONTRACT), "style_prompt_contract": asset_record(STYLE_CONTRACT), "illustration_prompt_contract": asset_record(ILLUSTRATION_CONTRACT), "prompt_contracts_applied": {"watercolor_style": False, "illustration": True}, "inputs": [], "input_roles": [], "framing": args.framing, "seed": args.seed, "steps": args.steps, "size": [args.size, args.size], "true_cfg_scale": 4.0, "guidance_scale": 1.0, "negative_prompt": " ", "prompt": prompt, "prompt_word_count": len(prompt.split()), "output": asset_record(output), "elapsed_seconds": round(time.monotonic() - started, 2), } result_record.write_text(json.dumps(record, ensure_ascii=False, indent=2) + "\n", encoding="utf-8") print(json.dumps({"output": str(output), "result_record": str(result_record)}, ensure_ascii=False)) if __name__ == "__main__": main()