app.py

import os
import torch
import gradio as gr

from diffusers import AutoencoderTiny
from torchvision.transforms.functional import to_pil_image, center_crop, resize, to_tensor

device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'
vae = None


def get_model():
    global vae

    if vae is None:
        model_id = "madebyollin/taesd"
        vae = AutoencoderTiny.from_pretrained(model_id, safetensors=True).to(device)
    return vae


@torch.no_grad()
def encode(image):
    vae = get_model()

    DIM = [512]
    processed = center_crop(resize(image, DIM), DIM)
    tensor = to_tensor(processed).unsqueeze(0).to(device)
    latents = vae.encoder(tensor)
    scaled = vae.scale_latents(latents).mul_(255).round_().byte()
    return to_pil_image(scaled[0])


path = 'https://huggingface.co/buckets/ciCic/demo-purposes/resolve/images'
astronaut = f"{path}/6.png"


def app():
    return gr.Interface(encode,
                        gr.Image(type="pil",
                                 label='512x512',
                                 value=astronaut),
                        gr.Image(type="pil",
                                 image_mode="RGBA",
                                 label='64x64',
                                 height=256,
                                 width=256
                                 ),
                        examples=[
                            astronaut,
                            f"{path}/7.png",
                            f"{path}/34.png"
                        ], flagging_mode='never', title='Image Encoder')


if __name__ == "__main__":
    print("LAUNCHING")
    app().launch(server_name="0.0.0.0", server_port=7860, share=True)