KATok

Kakao's video-tokenizer paper, accepted to ECCV 2026 and announced by the company on 9 September. Conventional video VAEs compress at a fixed ratio regardless of how much is happening in a clip; KATok is a transformer VAE with an adaptive token selector, learned jointly with the latent tokens, that scores each token's content-richness as a keep-or-drop probability and discards the uninformative ones, so compression follows the content. Because dropping tokens misaligns the spatial grid that diffusion models expect, the paper adds cascaded and joint generation schemes to restore alignment, and reports state-of-the-art compression ratios; Kakao's release claims 3.2x faster generation and 6.9x less training compute for the downstream generator.