run_demo_inference.py

#!/usr/bin/env python3
"""Run ZoomLDM-CDM-brca demo inference and save visualization to demo_images/."""
from pathlib import Path

import numpy as np
import torch
from diffusers import DiffusionPipeline
from PIL import Image, ImageDraw


def main() -> None:
    repo = Path(__file__).resolve().parent
    demo_dir = repo / "demo_images"
    demo_dir.mkdir(exist_ok=True)

    pipe = DiffusionPipeline.from_pretrained(
        str(repo),
        custom_pipeline=str(repo / "pipeline.py"),
        trust_remote_code=True,
        local_files_only=True,
    ).to("cuda")

    magnification = torch.tensor([0], device="cuda")
    out = pipe(
        batch_size=1,
        magnification=magnification,
        num_inference_steps=50,
        guidance_scale=1.0,
    )

    # out.samples: (1, 512, 65) -> visualize first sample as a heatmap-like image
    sample = out.samples[0].detach().float().cpu().numpy()
    sample = (sample - sample.min()) / (sample.max() - sample.min() + 1e-8)
    sample_u8 = (sample * 255).astype(np.uint8)  # (512, 65)
    img = Image.fromarray(sample_u8, mode="L").resize((520, 512), Image.Resampling.NEAREST).convert("RGB")
    img.save(demo_dir / "output.jpeg")

    # Simple input card to make widget context explicit
    card = Image.new("RGB", (520, 512), color=(245, 245, 245))
    draw = ImageDraw.Draw(card)
    draw.text((24, 24), "ZoomLDM-CDM-brca", fill=(20, 20, 20))
    draw.text((24, 64), "Input: magnification class = 0", fill=(40, 40, 40))
    draw.text((24, 100), "Output: sampled conditioning tensor visualization", fill=(40, 40, 40))
    card.save(demo_dir / "input.jpeg")

    print(f"Saved {demo_dir / 'input.jpeg'}")
    print(f"Saved {demo_dir / 'output.jpeg'}")


if __name__ == "__main__":
    main()