KVAE 1.0: Image tokenizer

KVAE-Image 1.0 is the image tokenizer from KVAE 1.0, a family of image and video tokenizers designed for latent diffusion models. It compresses RGB images into compact continuous latent representations and reconstructs them with high fidelity.

Model zoo

Inference

Run from the KVAE source repository root. The reference environment uses Python 3.11, PyTorch 2.8.0, and CUDA 12.8.

pip install -r requirements.txt

import torch

from data import read_image
from kvae import KVAEImage

device = torch.device("cuda:0")
dtype = torch.bfloat16

model = KVAEImage.from_pretrained("kandinskylab/KVAE-2D-1.0").eval().to(device=device, dtype=dtype)
image = read_image("assets/image_test/0002.png").unsqueeze(0)
image = image.to(device=device, dtype=dtype)

with torch.no_grad():
    latent = model.encode(image).latent_dist.mode()
    reconstruction = model.decode(latent).clip(-1, 1)

Evaluation

Reconstruction was evaluated on the validation splits of ImageNet-256 and DIV2K. All compared models use 8 x 8 spatial compression and 16 latent channels.

Show reconstruction and DiT training figures

Reconstruction comparison

Columns from left to right: original image, KVAE-Image 1.0 reconstruction, and FLUX VAE reconstruction.

DiT training metrics

CLIP score and FID during DiT training: blue is DiT + FLUX; green and red are two versions of DiT + KVAE-Image 1.0.

Citation

@misc{kvae_1_2025,
  author       = {Kirill Chernyshev, Andrey Shutkin, Ilia Vasiliev,
                  Denis Parkhomenko, Ivan Kirillov,
                  Dmitrii Mikhailov, Denis Dimitrov},
  title        = {KVAE 1.0: Image and Video Tokenizers for Image and Video Generation Models},
  howpublished = {\url{https://github.com/kandinskylab/kvae}},
  year         = {2025}
}