demo_inference.py

#!/usr/bin/env python3
"""Generate a demo image with JiT-H-32."""

from pathlib import Path

import torch
from diffusers import DiffusionPipeline, FlowMatchHeunDiscreteScheduler

REPO_ROOT = Path(__file__).resolve().parent
MODEL_DIR = REPO_ROOT / "JiT-H-32"
OUTPUT_PATH = REPO_ROOT / "demo.png"


def main() -> None:
    pipe = DiffusionPipeline.from_pretrained(
        str(MODEL_DIR),
        custom_pipeline=str(MODEL_DIR / "pipeline.py"),
        trust_remote_code=True,
        torch_dtype=torch.bfloat16,
    )
    pipe.scheduler = FlowMatchHeunDiscreteScheduler.from_config(pipe.scheduler.config, shift=4.0)
    pipe.to("cuda")
    pipe.set_progress_bar_config(disable=False)

    print(pipe.id2label[207])
    print(pipe.get_label_ids("golden retriever"))

    generator = torch.Generator(device="cuda").manual_seed(42)
    image = pipe(
        class_labels="golden retriever",
        num_inference_steps=50,
        guidance_scale=2.3,
        generator=generator,
    ).images[0]
    image.save(OUTPUT_PATH)
    print(f"Saved demo image to {OUTPUT_PATH}")


if __name__ == "__main__":
    main()