Kandinsky WM 1.0 Autonomous Driving
This 2B world-model research checkpoint adapts Kandinsky 5.0 Video Lite to autonomous driving. It continues an initial image into a five-second scene.
Research release
Domain training changes the video transformer
The family uses domain adaptation followed by GRPO reinforcement learning against a physical-plausibility reward. The three releases have different DiT weights but identical VAE and text-encoder components.kandinskylab/Kandinsky-WM-1.0-I2V-5s-AV — model card and files
Use generated scenes as research data
Practical guidance. A video of a plausible interaction is not a calibrated physical simulator or a robot-control policy. Check the scene dynamics and the intended downstream task before using synthetic examples.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | kandinskylab/Kandinsky-WM-1.0-I2V-5s-AV |
| Base family | Kandinsky 5.0 Video Litekandinskylab/Kandinsky-WM-1.0-I2V-5s-AV — model card and files |
| Output | 121 frames at 768 × 512kandinskylab/Kandinsky-WM-1.0-I2V-5s-AV — model card and files |
| Weight license | MITkandinskylab/Kandinsky-WM-1.0-I2V-5s-AV — model card and files |
Related models
- Kandinsky 5.0 I2V Lite SFT 5s
Kandinsky Lab · Kandinsky 5.0 Video
A 2B Lite model that turns an image and motion description into a five-second clip.
Launch date not confirmedDownloadable