Forbu14/imagenet-1k-latent

Better latent: I advise you to use another dataset https://huggingface.co/datasets/cloneofsimo/imagenet.int8 which is already compressed (5Go only) and use a better latent model (SDXL)

This dataset is the latent representation of the imagenet dataset using the stability VAE stabilityai/sd-vae-ft-ema.

Every image_latent is of shape (4, 32, 32).

If you want to retrieve the original image you have to use the model used to create the latent image :

vae_model = "stabilityai/sd-vae-ft-ema"
vae = AutoencoderKL.from_pretrained(vae_model)
vae.eval()

The images have been encoded using :

images = [DEFAULT_TRANSFORM(image.convert("RGB")) for image in examples["image"]]
images = torch.stack(images)
images = vaeprocess.preprocess(images)
images = images.to(device="cuda", dtype=torch.float)
with torch.no_grad():
    latents = vae.encode(images).latent_dist.sample()

With DEFAULT_TRANSFORM being :

DEFAULT_IMAGE_SIZE = 256

DEFAULT_TRANSFORM = transforms.Compose(
    [
        transforms.Resize((DEFAULT_IMAGE_SIZE, DEFAULT_IMAGE_SIZE)),
        transforms.ToTensor(),
    ]
)

The images can be decoded using :

import datasets

latent_dataset = datasets.load_dataset(
            "Forbu14/imagenet-1k-latent"
        )

latent = torch.tensor(latent_dataset["train"][0]["latents"])
image = vae.decode(latent).sample