Kandinsky LabKandinsky 4
Kandinsky 4.0 V2A
The video-to-audio generator fine-tunes Riffusion's spectrogram diffusion model. Its modified conditioning uses video frames, with CogVLM2 providing visual and text representations.
Generate a spectrogram before the waveform
A diffusion U-Net predicts the spectrogram. Griffin–Lim converts that representation to sound. This is a distinct audio generator, not the audio track from the input video.ai-forever/kandinsky-4-v2a — model card and files
Read the derivative and upstream terms together
The Riffusion card states use restrictions and conditions for redistribution. The Kandinsky card's Apache label should not be read as removing the upstream terms.Riffusion v1: upstream model card and license
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ai-forever/kandinsky-4-v2a |
| Base model | riffusion/riffusion-model-v1ai-forever/kandinsky-4-v2a — model card and files |
| Kandinsky card license | Apache 2.0ai-forever/kandinsky-4-v2a — model card and files |
| Upstream Riffusion license | CreativeML OpenRAIL-MRiffusion v1: upstream model card and license |
Related models
- Kandinsky 4.0 T2V Flash
Kandinsky Lab · Kandinsky 4
A distilled Kandinsky 4 model for twelve-second 480p video, with a CogVideoX VAE and T5 encoder.
Launch date not confirmedDownloadable