demo_inference.py

#!/usr/bin/env python3
"""Generate a demo image with DiCo 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 = {
    "xl": {
        "dir": REPO_ROOT / "DiCo-XL-256",
        "class_label": "golden retriever",
        "num_inference_steps": 250,
        "guidance_scale": 1.4,
        "seed": 0,
    },
    "s": {
        "dir": REPO_ROOT / "DiCo-S-256",
        "class_label": "golden retriever",
        "num_inference_steps": 250,
        "guidance_scale": 1.0,
        "seed": 0,
    },
    "b": {
        "dir": REPO_ROOT / "DiCo-B-256",
        "class_label": "golden retriever",
        "num_inference_steps": 250,
        "guidance_scale": 1.0,
        "seed": 0,
    },
    "l": {
        "dir": REPO_ROOT / "DiCo-L-256",
        "class_label": "golden retriever",
        "num_inference_steps": 250,
        "guidance_scale": 1.0,
        "seed": 0,
    },
}


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Run DiCo demo inference.")
    parser.add_argument(
        "--variant",
        choices=sorted(VARIANTS),
        default="xl",
        help="Checkpoint variant to sample (default: xl).",
    )
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    settings = VARIANTS[args.variant]
    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"[{args.variant}] {settings['class_label']} -> {pipe.get_label_ids(settings['class_label'])}")

    generator = torch.Generator(device="cuda").manual_seed(settings["seed"])
    image = pipe(
        class_labels=settings["class_label"],
        height=256,
        width=256,
        num_inference_steps=settings["num_inference_steps"],
        guidance_scale=settings["guidance_scale"],
        generator=generator,
    ).images[0]
    image.save(output_path)
    print(f"Saved demo image to {output_path}")


if __name__ == "__main__":
    main()