#!/usr/bin/env python3
from diffusers import UNetUnconditionalModel, DDIMScheduler, VQModel
import torch
import PIL.Image
import numpy as np
import tqdm

# load all models
unet = UNetUnconditionalModel.from_pretrained("./", subfolder="unet")
vqvae = VQModel.from_pretrained("./", subfolder="vqvae")
scheduler = DDIMScheduler.from_config("./", subfolder="scheduler")

# set to cuda
torch_device = "cuda" if torch.cuda.is_available() else "cpu"

unet.to(torch_device)
vqvae.to(torch_device)

# generate gaussian noise to be decoded
generator = torch.manual_seed(0)
noise = torch.randn(
    (1, unet.in_channels, unet.image_size, unet.image_size),
    generator=generator,
).to(torch_device)

# set inference steps for DDIM
scheduler.set_timesteps(num_inference_steps=50)

image = noise
for t in tqdm.tqdm(scheduler.timesteps):
    # predict noise residual of previous image
    with torch.no_grad():
        residual = unet(image, t)["sample"]

    # compute previous image x_t according to DDIM formula
    prev_image = scheduler.step(residual, t, image, eta=0.0)["prev_sample"]

    # x_t-1 -> x_t
    image = prev_image

# decode image with vae
with torch.no_grad():
    image = vqvae.decode(image)

# process image
image_processed = image.cpu().permute(0, 2, 3, 1)
image_processed = (image_processed + 1.0) * 127.5
image_processed = image_processed.numpy().astype(np.uint8)
image_pil = PIL.Image.fromarray(image_processed[0])

image_pil.save("generated_image.png")