app.py

import gradio as gr
import os

import torch
from diffusers import AutoencoderTiny
from torchvision.transforms.functional import to_pil_image, to_tensor

device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'

model_id = "madebyollin/taesd"
vae = AutoencoderTiny.from_pretrained(model_id, safetensors=True).to(device)


@torch.no_grad()
def decode(image):
    t = to_tensor(image).unsqueeze(0).to(device)
    unscaled_t = vae.unscale_latents(t)
    reconstructed = vae.decoder(unscaled_t).clamp(0, 1)
    return to_pil_image(reconstructed[0])


astronaut = os.path.join(os.path.dirname(__file__), "images/21.encoded.png")


def app():
    return gr.Interface(decode,
                        gr.Image(type="pil",
                                 image_mode="RGBA",
                                 mirror_webcam=False,
                                 label='64x64',
                                 value=astronaut),
                        gr.Image(type="pil",
                                 image_mode="RGB",
                                 label='512x512',
                                 show_share_button=True,
                                 height=512,
                                 width=512
                                 ),
                        css=".upload-container {width: 256px !important; height:256px !important} ",
                        examples=[
                            astronaut,
                            os.path.join(os.path.dirname(__file__), "images/18.encoded.png"),
                            os.path.join(os.path.dirname(__file__), "images/20.encoded.png")
                        ], allow_flagging='never', title='Latents Decoder')


if __name__ == "__main__":
    app().launch()