app.py

import gradio as gr
import torch
import torchvision.transforms as T
from PIL import Image
from diffusers import AsymmetricAutoencoderKL
import spaces

MODEL_ID = "AiArtLab/sdxs-1b"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32


# -------------------------
# Load VAE
# -------------------------
def load_vae(model_id=MODEL_ID):
    for attempt in (None, "vae"):
        try:
            if attempt is None:
                vae = AsymmetricAutoencoderKL.from_pretrained(
                    model_id,
                    torch_dtype=DTYPE
                )
            else:
                vae = AsymmetricAutoencoderKL.from_pretrained(
                    model_id,
                    subfolder=attempt,
                    torch_dtype=DTYPE
                )

            vae = vae.to(DEVICE)
            vae.eval()

            print("VAE loaded on", DEVICE)
            return vae

        except Exception as e:
            last_err = e

    raise RuntimeError(f"Failed to load VAE: {last_err}")


_vae = None


def get_vae():
    global _vae
    if _vae is None:
        _vae = load_vae()
    return _vae


# -------------------------
# Encode / Decode
# -------------------------
@spaces.GPU(duration=60)
def encode_decode(img: Image.Image):

    if img is None:
        raise gr.Error("Please upload an image")

    vae = get_vae()

    img = img.convert("RGB")

    tfm = T.Compose([
        T.ToTensor(),
        T.Normalize([0.5]*3, [0.5]*3),
    ])

    t = tfm(img).unsqueeze(0).to(DEVICE, dtype=DTYPE)

    print("Input tensor:", t.shape, t.dtype, t.device)

    with torch.no_grad():

        enc = vae.encode(t)
        lat = enc.latent_dist.sample()

        print("Latents:", lat.shape)

        dec = vae.decode(lat).sample

    x = (dec.clamp(-1, 1) + 1) * 127.5
    x = x.round().to(torch.uint8)

    x = x.squeeze(0).permute(1, 2, 0).cpu().numpy()

    out = Image.fromarray(x)

    print("Output size:", out.size)

    return out


# -------------------------
# UI
# -------------------------

with gr.Blocks(title="VAE True-to-Source Upscaler") as demo:
    gr.Markdown("# VAE-based 2x Upscaler\nSimple, blind, true-to-source upscaling without AI hallucinations.")
    
    with gr.Row():
        with gr.Column():
            input_img = gr.Image(type="pil", label="Upload Image")
            upscale_btn = gr.Button("Upscale x2", variant="primary")
        with gr.Column():
            output_img = gr.Image(type="pil", label="Result (x2)")

    # Привязка кнопки к функции
    upscale_btn.click(
        fn=encode_decode, 
        inputs=input_img, 
        outputs=output_img
    )


# -------------------------
# Launch
# -------------------------
if __name__ == "__main__":
    demo.launch()