K-VAE

KVAE 2.0 Image is a continuous image tokenizer designed for diffusion-based text-to-image generation. It uses 8×8 spatial compression with 32 latent channels and has approximately 0.2B parameters. In our reconstruction benchmarks, KVAE 2.0 Image outperforms the VAEs used in leading open-source text-to-image models in PSNR and SSIM. Its latent space also translates into stronger downstream generation: KVAE 2.0 Image is preferred by human side-by-side evaluation under a fixed text-to-image setup, which assesses prompt following, visual quality, and semantic consistency.

A continuous, full-band (48 kHz) audio autoencoder. It compresses raw waveforms into compact continuous latents and reconstructs them with high fidelity across speech, music, and general sound. The model is designed not only for faithful reconstruction, but as a latent space for generative models — in our internal text-to-audio pipeline, swapping the autoencoder for KVAE 1.0 Audio improves generation quality under a fixed generator.

A family of causal video tokenizers built for diffusion-based generative models. In our reconstruction benchmarks and side-by-side evaluations, the released models outperform the compared open‑source tokenizers on key measures of reconstruction and generation quality. Both variants use 4× temporal compression, with 8×8 or 16×16 spatial compression. A purely convolutional architecture and cached temporal context enable efficient chunked processing of long videos.