demo_inference.py

#!/usr/bin/env python3
"""Generate demo images for PixelDiT class-conditional checkpoints."""

from __future__ import annotations

import argparse
from pathlib import Path

import torch
from diffusers import DiffusionPipeline

REPO_ROOT = Path(__file__).resolve().parent

VARIANTS = {
    "256": {
        "dir": REPO_ROOT / "PixelDiT-XL-16-256",
        "height": 256,
        "width": 256,
        "num_inference_steps": 100,
        "guidance_scale": 3.25,
        "class_label": "golden retriever",
        "seed": 7,
    },
    "512": {
        "dir": REPO_ROOT / "PixelDiT-XL-16-512",
        "height": 512,
        "width": 512,
        "num_inference_steps": 100,
        "guidance_scale": 3.75,
        "class_label": "golden retriever",
        "seed": 7,
    },
}


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Run PixelDiT demo inference.")
    parser.add_argument(
        "--variant",
        choices=sorted(VARIANTS),
        default="256",
        help="Checkpoint resolution variant to sample.",
    )
    parser.add_argument(
        "--all",
        action="store_true",
        help="Generate demo.png for every supported variant.",
    )
    return parser.parse_args()


def run_variant(name: str) -> Path:
    settings = VARIANTS[name]
    model_dir = settings["dir"]
    output_path = model_dir / "demo.png"

    pipe = DiffusionPipeline.from_pretrained(
        str(model_dir),
        local_files_only=True,
        custom_pipeline=str(model_dir / "pipeline.py"),
        trust_remote_code=True,
        torch_dtype=torch.bfloat16,
    )
    pipe.to("cuda")

    print(f"[{name}] {settings['class_label']} -> {pipe.get_label_ids(settings['class_label'])}")
    print(f"[{name}] scheduler shift={pipe.scheduler.config.shift}")

    generator = torch.Generator(device="cuda").manual_seed(settings["seed"])
    image = pipe(
        class_labels=settings["class_label"],
        height=settings["height"],
        width=settings["width"],
        num_inference_steps=settings["num_inference_steps"],
        guidance_scale=settings["guidance_scale"],
        guidance_interval_min=0.1,
        guidance_interval_max=1.0,
        generator=generator,
    ).images[0]
    image.save(output_path)
    print(f"[{name}] Saved demo image to {output_path}")
    return output_path


def main() -> None:
    args = parse_args()
    if args.all:
        for name in VARIANTS:
            run_variant(name)
        return
    run_variant(args.variant)


if __name__ == "__main__":
    main()