| """Generate one planned person and ten expression samples directly into a TAR.""" |
| import argparse |
| import hashlib |
| import io |
| import json |
| from pathlib import Path |
| import tarfile |
| import time |
| from urllib.request import urlopen |
| from PIL import Image |
| import torch |
| from native_flux_inference import create_native_pipeline, generate_native_image, encode_text |
| from background_production import add_member |
| from character_pilot import metadata |
|
|
| def main(): |
| p=argparse.ArgumentParser();p.add_argument('--person',default='person-0253') |
| p.add_argument('--plan',type=Path);p.add_argument('--base-path',type=Path) |
| p.add_argument('--output',type=Path,default=Path('person-emotions.tar')) |
| p.add_argument('--revision',default='main',help='Dataset plan revision; model revisions are separately pinned') |
| p.add_argument('--no-compile',action='store_true');a=p.parse_args() |
| if a.plan:plan=json.loads(a.plan.read_text()) |
| else: |
| url=f'https://huggingface.co/datasets/laion/Vivarium-Greenscreen-Characters-FLUX-9B/resolve/{a.revision}/metadata/person-plan.json' |
| with urlopen(url,timeout=90) as f:plan=json.load(f) |
| person=next(x for x in plan if x['personality_id']==a.person) |
| pipe=create_native_pipeline('characters',base_path=a.base_path,compile_mode=None if a.no_compile else 'default') |
| prompts={e['id']:encode_text(pipe,e['prompt']) for e in person['expressions']} |
| reference=None;reference_sha='';a.output.parent.mkdir(parents=True,exist_ok=True) |
| temporary=a.output.with_name(a.output.name+'.partial') |
| with tarfile.open(temporary,'w',format=tarfile.USTAR_FORMAT) as tar: |
| for i,e in enumerate(person['expressions']): |
| torch.cuda.synchronize();start=time.perf_counter() |
| image=generate_native_image(pipe,prompt_embeds=prompts[e['id']],seed=e['seed'],steps=50,guidance=4., |
| negative_prompt=e['negative_prompt'],image=reference if i else None,width=832,height=1248)[0] |
| torch.cuda.synchronize();seconds=time.perf_counter()-start |
| b=io.BytesIO();image.save(b,format='PNG',compress_level=4);png=b.getvalue() |
| m=metadata(person,e,png,reference_sha,seconds,0.,a.output.name,'portable-inference') |
| m['inference']['compile_mode']=None if a.no_compile else 'default' |
| m['runtime']['positive_text_encoding_seconds']=None |
| m['runtime']['includes_compile_warmup']=not a.no_compile and i<2 |
| add_member(tar,e['id']+'.png',png);add_member(tar,e['id']+'.json',json.dumps(m,indent=2,ensure_ascii=False).encode()) |
| add_member(tar,e['id']+'.txt',e['prompt'].encode()) |
| if i==0:reference=image.resize((416,624),Image.Resampling.BICUBIC);reference_sha=hashlib.sha256(png).hexdigest() |
| print(e['id'],round(seconds,2),'seconds',flush=True) |
| temporary.replace(a.output) |
|
|
| if __name__=='__main__':main() |
|
|
