scripts/infer.py

"""Local or Hub inference for the published Diffusers checkpoint."""
import argparse
from pathlib import Path

import torch
from scripts.runtime import load_int8_pipeline
from scripts.editing_protocol import load_input, validate_seeds, DEFAULT_SEEDS, edit_dimensions


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument('--model', default='models/int8')
    ap.add_argument('--prompt', required=True)
    ap.add_argument('--input', nargs='*', default=[])
    ap.add_argument('--output', default='output.png')
    ap.add_argument('--seed', type=int, help='Default: 42 for generation, 1000042 for editing')
    ap.add_argument('--source-seed', type=int, help='Known source generation seed; reject editing noise replay')
    ap.add_argument('--width', type=int)
    ap.add_argument('--height', type=int)
    ap.add_argument('--resolution', type=int, default=1024,
                    help='Editing reference/target area scale, independent of explicit output width')
    ap.add_argument('--steps', type=int, help='Default: checkpoint schedule, otherwise 40')
    ap.add_argument('--vae-tiling', action='store_true',
                    help='Save VAE memory; may introduce the stripe artifacts documented in EDITING_UPDATE.md')
    args = ap.parse_args()
    if ((args.width is None) != (args.height is None)):
        ap.error('Specify both --width and --height, or neither')
    if (args.steps is not None and args.steps <= 0) or any(v <= 0 or v % 32 for v in
                             (args.width, args.height, args.resolution) if v is not None):
        ap.error('Positive steps and dimensions divisible by 32 are required')
    if Path(args.output).suffix.lower() != '.png':
        ap.error('Use a .png output to retain the RGBA channel')
    images = [load_input(path) for path in args.input]
    seed = args.seed if args.seed is not None else (DEFAULT_SEEDS[0] if images else 42)
    try:
        validate_seeds([{'source_seed': args.source_seed}] if images else [], [seed])
    except ValueError as exc:
        ap.error(str(exc))
    if args.width is None:
        args.width, args.height = (edit_dimensions(images[-1].size, args.resolution)
                                   if images else (args.resolution, args.resolution))
    if Path(args.model).exists():
        from scripts.benchmark import load_pipeline
        pipe = load_pipeline(args.model)
    else:
        pipe = load_int8_pipeline(args.model)
    sigmas = pipe.config.get('sample_sigmas')
    if sigmas is not None and args.steps is not None and args.steps != len(sigmas):
        ap.error(f'This checkpoint uses a fixed {len(sigmas)}-step schedule')
    steps = args.steps if args.steps is not None else len(sigmas) if sigmas is not None else 40
    if args.vae_tiling:
        pipe.vae.enable_tiling()
    with torch.inference_mode():
        image = pipe(prompt=args.prompt, image=images or None,
                     width=args.width, height=args.height, output_resolution=args.resolution,
                     num_inference_steps=steps, true_cfg_scale=1.0, use_kv_cache=True,
                     generator=torch.Generator('cpu').manual_seed(seed)).images[0]
    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
    image.save(args.output)
    print(f'{args.output} (seed={seed}, actual_size={image.size}, mode={image.mode})')


if __name__ == '__main__':
    main()