#!/usr/bin/env python3 """Apply BFS to Scenes A/B with direction-matched eye-level head crops. First run --prepare-reference to crop source pixels without generation or masks. Then run the default command to apply BFS to the reviewed Scene A/B identity results. """ from __future__ import annotations import argparse import json import re import time from pathlib import Path import hashlib import importlib.metadata import platform import sys from PIL import Image ASSETS = Path(__file__).resolve().parent.parent PROJECT_ROOT = ASSETS.parents[3] CACHE_DIR = PROJECT_ROOT / ".tmp" / "download" / "huggingface" / "hub" MODEL_ID = "Qwen/Qwen-Image-Edit-2511" 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 runtime_record() -> dict[str, object]: packages = {} for package in ("diffusers", "torch", "transformers", "accelerate"): try: packages[package] = importlib.metadata.version(package) except importlib.metadata.PackageNotFoundError: packages[package] = "not-installed" return { "python": sys.version.split()[0], "platform": platform.platform(), "packages": packages, } def square_canvas(path: Path, size: int) -> Image.Image: """Return an RGB, white-backed square canvas without distorting the input.""" with Image.open(path) as source: source = source.convert("RGBA") source.thumbnail((size, size), Image.Resampling.LANCZOS) canvas = Image.new("RGBA", (size, size), "white") offset = ((size - source.width) // 2, (size - source.height) // 2) canvas.alpha_composite(source, offset) return canvas.convert("RGB") SOURCES = { 'a': 'sec-05/p7-5-5-qwen-2511-cutout-identity-a-mira-audit-20260909-v1-size-1280x1280-seed-62294-steps-30.png', 'b': 'sec-05/p7-5-5-qwen-2511-mannequin-outfit-b-mira-stage3-v1-size-1280x1280-seed-62294-steps-20.png', 'c': 'sec-05/p7-5-5-qwen-2511-identity-c-mira-stage2-book-v1-size-1280x1280-seed-62294-steps-20.png', } DEFAULT_FACE = ASSETS / 'sec-02/p7-5-2-mira-head-qwen-image-bf16-front-v1-code-63ece7-seed-62294-steps-30-size-1280.png' REFERENCES = { 'a': ('sec-02/p7-5-2-qwen-2511-mira-torso-multiview-vertical-level-yaw-zero-native1280-v1-size-1280x1280-seed-62294-steps-4.png', (320, 10, 960, 650), False, 1280), 'b': ('sec-02/p7-5-2-qwen-2511-mira-torso-multiview-vertical-level-yaw-plus-45-native1024-v1-size-1024x1024-seed-62294-steps-4.png', (280, 10, 792, 522), False, 1024), } ZERO_REFERENCES = { 'a': ('sec-02/p7-5-2-qwen-2511-mira-torso-multiview-vertical-low-yaw-zero-lowzero-repeat-v1-size-1280x1280-seed-62295-steps-4.png', (400, 40, 912, 552), False, 1280), 'b': ('sec-02/p7-5-2-qwen-2511-mira-torso-multiview-vertical-level-yaw-zero-native1280-v1-size-1280x1280-seed-62294-steps-4.png', (320, 10, 960, 650), False, 1280), } LOW45_REFERENCES = { 'b': ('sec-02/p7-5-2-qwen-2511-mira-torso-multiview-vertical-low-yaw-plus-45-native1280-v1-size-1280x1280-seed-62294-steps-4.png', (360, 20, 1000, 660), False, 1280), } def crop_paths(scene: str, profile: str = 'direction-v2') -> tuple[Path, Path]: variant = {'direction-v2': 'level-matched-native-head-crop-v2', 'zero-v3': 'zero-view-native-head-crop-v3', 'low45-v4': 'low45-native-head-crop-v4'}[profile] stem = f'p7-5-5-bfs-{scene}-{variant}' return ASSETS / 'sec-05' / f'{stem}.png', ASSETS / 'sec-05' / f'{stem}-result.json' LORA_FILE = 'bfs_head_v5_2511_original.safetensors' LORA_PATH = ASSETS.parents[3] / '.tmp/download/artifacts/weight-mr2along-bfs-head-v5-2511' / LORA_FILE LORA_SHA256 = '5055ec271228d4767fd327efa59dbdea6f2a628d6d890101eb4fd5015cb1ac57' PROMPT = ( "head_swap: start with Picture 1 as the base image, keeping its lighting, " "environment, and background. remove the head from Picture 1 completely and " "replace it with the head from Picture 2, strictly preserving the hair, eye " "color, and nose structure of Picture 2. copy the eye direction, head rotation, " "and micro-expressions from Picture 1. high quality, sharp details, 4k" ) def main() -> None: parser = argparse.ArgumentParser(description=__doc__) parser.add_argument('--scenes', nargs='+', choices=('a', 'b'), default=['a', 'b']) parser.add_argument('--prepare-reference', action='store_true') parser.add_argument('--reference-profile', choices=('direction-v2', 'zero-v3', 'low45-v4'), default='direction-v2') parser.add_argument('--prompt', default=PROMPT) parser.add_argument('--steps', type=int, default=10) parser.add_argument('--seed', type=int, default=62294) parser.add_argument('--run-label', default='v2') parser.add_argument('--output-dir', type=Path, default=ASSETS / "sec-05") parser.add_argument('--dry-run', action='store_true') args = parser.parse_args() if args.steps < 1 or not args.prompt.strip(): parser.error('Positive steps and a nonempty prompt are required.') if not re.fullmatch(r'[A-Za-z0-9][A-Za-z0-9_-]*', args.run_label): parser.error('Use letters, digits, underscores and hyphens for --run-label.') references = {'direction-v2': REFERENCES, 'zero-v3': ZERO_REFERENCES, 'low45-v4': LOW45_REFERENCES}[args.reference_profile] if any(scene not in references for scene in args.scenes): parser.error('The selected reference profile only supports: ' + ', '.join(references)) if args.prepare_reference: from PIL import Image, ImageOps for view in dict.fromkeys(args.scenes): source_name, crop_box, mirror, source_size = references[view] source = ASSETS / source_name crop_path, crop_record = crop_paths(view, args.reference_profile) for path in (crop_path, crop_record): if path.exists(): raise FileExistsError(path) with Image.open(source) as original: if original.size != (source_size, source_size): raise ValueError(f'Expected native {source_size}px torso: {source}') cropped = original.crop(crop_box) if mirror: cropped = ImageOps.mirror(cropped) cropped.save(crop_path) crop_record.write_text(json.dumps(dict( operation='pixel crop with optional horizontal flip; no resizing or AI generation', horizontal_flip=mirror, input=str(source), input_sha256=sha256(source), crop_box=list(crop_box), output=str(crop_path), size=list(cropped.size), output_sha256=sha256(crop_path), source_code_sha256=sha256(Path(__file__)), ), indent=2) + '\n') print(crop_path) return if not LORA_PATH.is_file() or sha256(LORA_PATH) != LORA_SHA256: raise ValueError(f'Missing or mismatched BFS weights: {LORA_PATH}') plans = [] for experiment in [{'direction-v2': 'level-matched', 'zero-v3': 'zero-view', 'low45-v4': 'low45'}[args.reference_profile]]: for scene in dict.fromkeys(args.scenes): source = ASSETS / SOURCES[scene] reference, crop_record = crop_paths(scene, args.reference_profile) weight = 1.0 for path in (source, reference): if not path.is_file(): raise FileNotFoundError(path) variant = f'{experiment}-native-headcrop-weight10' stem = (f'p7-5-5-qwen-2511-bfs-reviewed-{scene}-mira-{variant}-{args.run_label}' f'-size-1280x1280-seed-{args.seed}-steps-{args.steps}') output = args.output_dir.resolve() / f'{stem}.png' result = output.with_name(f'{stem}-result.json') for path in (output, result): if path.exists(): raise FileExistsError(f'Refusing to overwrite: {path}') plans.append(dict(scene=scene, experiment=experiment, reference_profile=args.reference_profile, crop_provenance=json.loads(crop_record.read_text()), lora_weight=weight, inputs=[ dict(role='Picture 1: reviewed pre-BFS identity result', path=str(source), sha256=sha256(source)), dict(role='Picture 2: native head crop from direction-matched torso reference', path=str(reference), sha256=sha256(reference)), ], output=str(output), result=str(result))) settings = dict(model=MODEL_ID, prompt=args.prompt.strip(), steps=args.steps, seed=args.seed, size=[1280, 1280], true_cfg_scale=4.0, guidance_scale=1.0, generator_device='cpu', dtype='bfloat16', mask=None, lora=dict(repository='mr2along/BFS', revision='d9ffba9012dfa1b299a4294572791c3275ae6ae4', file=LORA_FILE, sha256=LORA_SHA256)) if args.dry_run: print(json.dumps(dict(status='planned', settings=settings, plans=plans), indent=2)) return import torch from diffusers import QwenImageEditPlusPipeline if not torch.cuda.is_available(): raise RuntimeError('A local CUDA GPU is required.') pipeline = QwenImageEditPlusPipeline.from_pretrained( MODEL_ID, torch_dtype=torch.bfloat16, cache_dir=CACHE_DIR, local_files_only=True, ) pipeline.load_lora_weights(LORA_PATH.parent, weight_name=LORA_FILE, adapter_name='bfs_head_v5', local_files_only=True) pipeline.enable_attention_slicing('max') pipeline.enable_sequential_cpu_offload() args.output_dir.mkdir(parents=True, exist_ok=True) provenance = dict(source_code_sha256=sha256(Path(__file__)), runtime=runtime_record(), cuda_device=torch.cuda.get_device_name(0), offload='sequential CPU offload', preprocessing='aspect-preserving resize on white 1280x1280 canvases') for plan in plans: for item in plan['inputs']: if sha256(Path(item['path'])) != item['sha256']: raise ValueError(f"Input changed after preflight: {item['path']}") pipeline.set_adapters('bfs_head_v5', adapter_weights=plan['lora_weight']) print(f"Starting experiment {plan['experiment']} scene {plan['scene']}", flush=True) started = time.monotonic() image = pipeline( image=[square_canvas(Path(item['path']), 1280) for item in plan['inputs']], prompt=settings['prompt'], height=1280, width=1280, generator=torch.Generator(device='cpu').manual_seed(args.seed), true_cfg_scale=4.0, negative_prompt=' ', num_inference_steps=args.steps, guidance_scale=1.0, num_images_per_prompt=1, ).images[0] with Path(plan['output']).open('xb') as stream: image.save(stream, format='PNG') record = dict(plan, **settings, **provenance, status='generated', stage='bfs_head_crop_reference_comparison', output_sha256=sha256(Path(plan['output'])), elapsed_seconds=round(time.monotonic()-started, 2)) with Path(plan['result']).open('x', encoding='utf-8') as stream: json.dump(record, stream, ensure_ascii=False, indent=2) stream.write('\n') print(json.dumps({'scene': plan['scene'], 'output': plan['output'], 'result': plan['result']}), flush=True) if __name__ == '__main__': main()