Sber open-sources KVAE-2.0 video tokenizers
Sber has released open weights and code for KVAE-2.0, two video tokenizers it says beat Wan 2.2 and HunyuanVideo 1.5 on reconstruction metrics and generation quality.
Sber's Kandinsky team announced KVAE-2.0 on Habr. It follows KVAE-1.0, which came out a few months earlier, and has two models. One keeps the earlier 4x8x8 compression: 4x along the time axis and 8x along height and width. The other uses a more aggressive 4x16x16 format, the same move Wan and HunyuanVideo made in their latest versions. Code and weights are publicly available.
The architecture is still a causal convolutional autoencoder, with three main changes:
- Encoder bottleneck fixed. In KVAE-1.0, channel counts grew by less than 2x at each downsampling layer, so some information was lost.
- RMSNorm instead of GroupNorm. Normalization now runs per frame, not across 16–17-frame groups. Users can shorten the processed segment at higher resolutions to avoid running out of memory.
- More capacity in the decoder. In the 4x16x16 model, the decoder has 5.3 times as many weights as the encoder, up from 1.3 in the 4x8x8 models. The decoder is also fine-tuned with a discriminator to keep high-frequency detail.
The team also expanded the training data and filtered it for high-motion video. It raised the chance of sampling a still image into a batch from 0.2 to 0.3.
On the MCL-JCV and BVI-DVC datasets, Sber reports better PSNR, SSIM and LPIPS scores. The 4x8x8 model was tested against Wan 2.1 and HunyuanVideo 1.0, and the 4x16x16 model against Wan 2.2 and HunyuanVideo 1.5. Sber also ran side-by-side human evaluations of videos generated with each tokenizer. It reports that a latent-space metric it calls CDS correlates with generation quality (Pearson 0.82).
Sber ran all of these tests itself, and the excerpt doesn't include the full numerical results.