Kandinsky LabKandinsky S2V
Kandinsky S2V 1.0 Lite Distilled
The model generates a speaking avatar from an image and recorded speech. It animates the figure as well as the head. The speech is an input condition; this model does not create the audio.
Generate speech-aligned motion
Output runs at 25 frames per second. Long speech is processed in five-second chunks. The Lite release documents SD output at the reference image's aspect ratio.kandinskylab/Kandinsky-S2V-1.0-Lite-distill — model card and files
Use the correct conditioning components
The pipeline uses Qwen2.5-Omni-7B, CLIP and a HunyuanVideo VAE. These upstream components are credited separately from Kandinsky's trained video transformer.kandinskylab/Kandinsky-S2V-1.0-Lite-distill — model card and files
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | kandinskylab/Kandinsky-S2V-1.0-Lite-distillkandinskylab/Kandinsky-S2V-1.0-Lite-distill-Diffusers |
| Generative DiT | 2.3B parameterskandinskylab/Kandinsky-S2V-1.0-Lite-distill — model card and files |
| Sampling | Pi-Flow, 16 stepskandinskylab/Kandinsky-S2V-1.0-Lite-distill — model card and files |
| Weight license | MITkandinskylab/Kandinsky-S2V-1.0-Lite-distill — model card and files |