TL;DR
This page aggregates structured data for CADRE, the open-source video autoencoder family released by Paris-based lab Kyutai on October 23, 2026: model counts and sizes, per-model latency and frame rates on a single H100, motion and reconstruction metrics, training data volume, and licenses. All figures are current as of October 24, 2026.
Data Overview
Sources: Kyutai's official paper "Rethinking Video Autoencoders for Interactive World Models" (arXiv 2610.30129, 2026), the Kyutai papers page, Sakhanda Wire and AiKraft reproductions of MarkTechPost coverage (October 23-24, 2026), and a Japanese summary by my-best (October 24, 2026). Scope and method: CADRE is a family of 7 causal diffusion representation autoencoders spanning different spatio-temporal compressions; members differ only in a small bottleneck on frozen features and share a diffusion decoder self-distilled to one or a few steps. Latency and frame-rate figures are Kyutai's own measurements on a single NVIDIA H100 (bf16, batch 1, 288x512 resolution, 1 decoding step), with consistent numbers reproduced by third-party reports. For FVD lower is better (motion quality), for ARR higher is better (action following), and for PSNR higher is better (reconstruction quality).
Core Data
| model | type | latency_ms | max_fps | license |
|---|---|---|---|---|
| CADRE-2x32x32 | CADRE causal diffusion representation autoencoder | 185 | 23.4 | code MIT, weights CC BY-NC 4.0 |
| CADRE-2x16x16 | CADRE member with lower spatial compression | not separately listed | not separately listed | code MIT, weights CC BY-NC 4.0 |
| Wan 2.2 VAE | text-to-video reuse autoencoder | 317 | 34.2 | set by each model |
| LTX-2 VAE | text-to-video reuse autoencoder | 492 | 87.4 | set by each model |
| Cosmos CV8x8x8 | text-to-video reuse autoencoder | 605 | 39.0 | set by each model |
| DC-AE | frame-by-frame image autoencoder | 118 | not separately listed | set by each model |
| metric | CADRE-2x32x32 | CADRE-2x16x16 | Wan 2.2 | LTX-2 | Cosmos |
|---|---|---|---|---|---|
| Driving FVD (lower is better) | 171.3 | 162.6 | 599.7 | 913.3 | 452.5 |
| Robot manipulation ARR (higher is better) | 0.45 | not separately listed | 0.85 | 0.25 | 0.78 |
| Reconstruction PSNR (higher is better) | 22.86 | not separately listed | 35.02 | 28.24 | 31.39 |
Field Notes
model: autoencoder name; in CADRE naming, the first digit is the temporal compression factor and the following digits are the spatial compression factor (for example, 2x32x32 means 2x temporal and 32x spatial compression).type: family and use; CADRE members are purpose-built for interactive world models, while the baselines are text-to-video or frame-by-frame compression components.latency_ms: per-frame action latency in milliseconds as self-measured by Kyutai on one H100 with 1 decoding step; the threshold at which humans begin to notice action lag is around 230 ms.max_fps: maximum generation frame rate on the same H100 setup, per Kyutai's own measurements.license: CADRE code is MIT licensed and its weights use CC BY-NC 4.0, which excludes commercial use.- FVD stands for Frechet Video Distance; ARR stands for Action Regression Rate; PSNR stands for peak signal-to-noise ratio.
Sources
- Kyutai official paper (arXiv 2610.30129): https://arxiv.org/pdf/2610.30129
- Kyutai papers page: https://kyutai.org/papers/
- Sakhanda Wire reproduction of MarkTechPost (October 23, 2026): https://wire.sakhanda.co/en/r/kyutai-cadre-video-autoencoders-that-cut-world-model-lag-below-230-ms
- AiKraft reproduction (October 24, 2026): https://aikraft.ru/news/kyutai-cadre-video-autoencoders-that-cut-world-model-lag-below-230-ms
- my-best Japanese summary (October 24, 2026): https://my-best.ai/kyutai-cadre-video-autoencoder-230ms/