KVAE-2D-2.0

<div align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/kandinskylab/KVAE-2D-2.0/resolve/main/assets/kvae-white-2.0.png" > <source media="(prefers-color-scheme: light)" srcset="https://huggingface.co/kandinskylab/KVAE-2D-2.0/resolve/main/assets/kvae-black-2.0.png" > <img src="https://huggingface.co/kandinskylab/KVAE-2D-2.0/resolve/main/assets/kvae-black-2.0.png" alt="KVAE 2.0 logo" width="600" > </picture>

<a href="https://github.com/kandinskylab/kvae">GitHub</a> | <a href="https://huggingface.co/papers/2608.05798">Technical report</a>

</div>

KVAE 2.0: Image tokenizer

KVAE-Image 2.0 is the image tokenizer from KVAE 2.0, a family of image and video tokenizers designed to provide latent representations for diffusion models. It compresses RGB images by 8 x 8 into continuous 32-channel latents and reconstructs them with high fidelity.

Model zoo

Versions

A new revision v2.1 is available and main points to it (from_pretrained(..., revision="v2.0") for the initial release); reconstruction quality on COCO val2017 (5000 images, native resolution, center crop to a multiple of 32, bf16):

In a side-by-side human evaluation of text-to-image generations, KVAE-Image 2.1 is preferred over 2.0 with an overall win rate of 54/46, driven by prompt following (56/44) and semantic quality (56/44), with visual quality on par (51/49).

Note: v2.0 and v2.1 have different encoders, so their latent spaces are not compatible.

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
pip install --editable .

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-2.0").eval().to(device=device, dtype=dtype)

image = read_image("path/to/image.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)

Citation

@misc{kvae2026,
  title         = {KVAE: Family of Tokenizers for Multimodal Generative Models},
  author        = {Andrey Shutkin, Denis Parkhomenko, Ivan Kirillov,
                   Kirill Chernyshev, Kirill Malakhov, Ilia Vasiliev,
                   Ilia Trushkin, Valeriya Kobenko, David Chikovani,
                   Alexander Ivanov, Azat Saginbaev, Egor Silvestrov,
                   Ivan Mikheev, Konstantin Zakharov},
  year          = {2026},
  eprint        = {2608.05798},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.05798}
}