Kandinsky Lab
Kandinsky Lab publishes the Kandinsky model family and related research. Its history runs from Russian text-to-image models to synchronized video and sound. Earlier releases credit Sber AI and AIRI; the current lab presents itself as a research team.
Models & tools
- Kandinsky 6.0 Image Pro
Kandinsky Lab · Kandinsky Image
A mixture-of-experts image model for text-to-image generation and instruction-based editing.
Launch date Provider product - Kandinsky 4.1 Image
Kandinsky Lab · Kandinsky Image
An image-generation model that replaces U-Net with a diffusion transformer and uses selected artistic training data.
Launch date Access not confirmed - Kandinsky 2.0
Kandinsky Lab · Kandinsky 2
An image-generation model with latent diffusion and two multilingual text encoders.
Launch date not confirmedDownloadable - Kandinsky 2.0 Inpainting
Kandinsky Lab · Kandinsky 2
A Kandinsky 2.0 checkpoint that fills masked image regions from a text prompt.
Launch date not confirmedDownloadable - Kandinsky 2.1
Kandinsky Lab · Kandinsky 2
An image model that combines an image prior and latent decoder with mixed text and image conditions.
Launch date not confirmedDownloadable - Kandinsky 2.1 Image Prior
Kandinsky Lab · Kandinsky 2
A trained diffusion prior that supplies image embeddings to the Kandinsky 2.1 decoder.
Launch date not confirmedDownloadable - Kandinsky 2.1 Inpainting
Kandinsky Lab · Kandinsky 2
A Kandinsky 2.1 checkpoint that fills masked image regions from a text prompt.
Launch date not confirmedDownloadable - Kandinsky 2.2
Kandinsky Lab · Kandinsky 2
A 1024-pixel image model with a retrained CLIP-based prior and an adapted diffusion decoder.
Launch date not confirmedDownloadable - Kandinsky 2.2 ControlNet Depth
Kandinsky Lab · Kandinsky control models
A Kandinsky 2.2 adaptation that uses depth maps to guide image generation.
Launch date not confirmedDownloadable - Kandinsky 2.2 Image Prior
Kandinsky Lab · Kandinsky 2
A trained prior that converts text conditions into CLIP image embeddings for Kandinsky 2.2.
Launch date not confirmedDownloadable - Kandinsky 2.2 Inpainting
Kandinsky Lab · Kandinsky 2
An inpainting decoder that combines an input image, mask and Kandinsky 2.2 prior condition.
Launch date not confirmedDownloadable - Kandinsky 3 ControlNet HED
Kandinsky Lab · Kandinsky control models
A Kandinsky 3 ControlNet research model conditioned on HED image edges.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 3 ControlNet v2 Depth
Kandinsky Lab · Kandinsky control models
A Kandinsky 3 ControlNet v2 model for depth-conditioned image generation.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 3 ControlNet v2 HED
Kandinsky Lab · Kandinsky control models
A Kandinsky 3 ControlNet v2 model for HED edge-conditioned image generation.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 3 ControlNet v2 Scribble
Kandinsky Lab · Kandinsky control models
A Kandinsky 3 ControlNet v2 model for scribble-conditioned image generation.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 3 Flash
Kandinsky Lab · Kandinsky 3
A distilled four-step Kandinsky 3 image model also used as the 3.1 release's refiner.
Launch date not confirmedDownloadable - Kandinsky 3 IP-Adapter
Kandinsky Lab · Kandinsky control models
A trained image-prompt adapter that conditions Kandinsky 3 generation on a reference image.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 3.0
Kandinsky Lab · Kandinsky 3
An image diffusion model with a larger text encoder and U-Net, trained with attention to Russian cultural subjects.
Launch date not confirmedDownloadable - Kandinsky 3.0 Inpainting
Kandinsky Lab · Kandinsky 3
A Kandinsky 3.0 derivative trained to fill masked image regions.
Launch date not confirmedDownloadable - Kandinsky 3.1 Inpainting
Kandinsky Lab · Kandinsky 3
An updated inpainting model trained with additional object-detection data.
Launch date not confirmedDownloadable - 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 - Kandinsky 4.0 V2A
Kandinsky Lab · Kandinsky 4
A Riffusion-derived model that generates audio from video frames using CogVLM2 representations.
Launch date not confirmedDownloadable - Kandinsky 4.0 Video
Kandinsky Lab · Kandinsky 4
A Kandinsky video model for text and image conditioning whose full T2V and I2V public releases remain forthcoming.
Launch date not confirmedAccess not confirmed - Kandinsky 5.0 I2V Lite Arc left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Arc left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Lite Arc right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Arc right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Lite Dolly in LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Dolly in LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Lite Dolly out LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Dolly out LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Lite Microwave left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Microwave left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Lite Microwave right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Microwave right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - 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 - Kandinsky 5.0 I2V Lite Truck left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Truck left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Lite Truck right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Lite Truck right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Pro Arc left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Pro Arc left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Pro Arc right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Pro Arc right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Pro Distilled 5s
Kandinsky Lab · Kandinsky 5.0 Video
A distilled 19B Pro model that animates a supplied image into a five-second clip.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Pro Microwave left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Pro Microwave left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Pro Microwave right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 I2V Pro Microwave right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 I2V Pro SFT 5s
Kandinsky Lab · Kandinsky 5.0 Video
A 19B Pro model that turns an image and motion description into a five-second clip.
Launch date not confirmedDownloadable - Kandinsky 5.0 Image Editing
Kandinsky Lab · Kandinsky 5.0 Image
A 6B image-editing model that changes an input picture according to a text instruction.
Launch date not confirmedDownloadable - Kandinsky 5.0 Image Editing Pretrain
Kandinsky Lab · Kandinsky 5.0 Image
A pretrained 6B image-editing checkpoint that accepts an image and a text instruction.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 5.0 Image Lite
Kandinsky Lab · Kandinsky 5.0 Image
A 6B text-to-image model for square and wide images, including Russian text and concepts.
Launch date not confirmedDownloadable - Kandinsky 5.0 Image Lite Pretrain
Kandinsky Lab · Kandinsky 5.0 Image
A pretrained 6B text-to-image model with square and wide output formats.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 5.0 T2V Lite Arc left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Arc left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Arc right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Arc right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Distilled 10s
Kandinsky Lab · Kandinsky 5.0 Video
A distilled 2B Lite model that generates ten-second video with reduced sampling work.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Distilled 5s
Kandinsky Lab · Kandinsky 5.0 Video
A distilled 2B Lite model that generates five-second video with reduced sampling work.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Dolly in LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Dolly in LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Dolly out LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Dolly out LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Microwave left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Microwave left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Microwave right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Microwave right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite No-CFG 10s
Kandinsky Lab · Kandinsky 5.0 Video
A 2B Lite text-to-video model with guidance distilled into its weights for ten-second clips.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite No-CFG 5s
Kandinsky Lab · Kandinsky 5.0 Video
A 2B Lite text-to-video model with guidance distilled into its weights for five-second clips.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Pretrain 10s
Kandinsky Lab · Kandinsky 5.0 Video
A pretrained 2B Lite model that generates ten-second video from text.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 5.0 T2V Lite Pretrain 5s
Kandinsky Lab · Kandinsky 5.0 Video
A pretrained 2B Lite model that generates five-second video from text.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 5.0 T2V Lite SFT 10s
Kandinsky Lab · Kandinsky 5.0 Video
A supervised-fine-tuned 2B Lite model for ten-second text-to-video generation.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite SFT 5s
Kandinsky Lab · Kandinsky 5.0 Video
A supervised-fine-tuned 2B Lite model for five-second text-to-video generation.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Truck left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Truck left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Lite Truck right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Lite Truck right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Pro Arc left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Pro Arc left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Pro Arc right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Pro Arc right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Pro Distilled 5s
Kandinsky Lab · Kandinsky 5.0 Video
A distilled 19B Pro model for five-second text-to-video generation.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Pro Microwave left LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Pro Microwave left LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Pro Microwave right LoRA
Kandinsky Lab · Kandinsky camera adapters
Kandinsky 5.0 T2V Pro Microwave right LoRA adds a specific learned motion to its matching video model.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Pro Pretrain 10s
Kandinsky Lab · Kandinsky 5.0 Video
A pretrained 19B Pro model for ten-second text-to-video generation.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 5.0 T2V Pro Pretrain 5s
Kandinsky Lab · Kandinsky 5.0 Video
A pretrained 19B Pro model for five-second text-to-video generation.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 5.0 T2V Pro SFT 10s
Kandinsky Lab · Kandinsky 5.0 Video
A supervised-fine-tuned 19B Pro model for ten-second text-to-video generation.
Launch date not confirmedDownloadable - Kandinsky 5.0 T2V Pro SFT 5s
Kandinsky Lab · Kandinsky 5.0 Video
A supervised-fine-tuned 19B Pro model for five-second text-to-video generation.
Launch date not confirmedDownloadable - Kandinsky 6.0 Video Lite
Kandinsky Lab · Kandinsky 6.0 Video
Kandinsky 6.0 Video Lite generates five-second video with synchronized sound.
Launch date not confirmedDownloadable - Kandinsky 6.0 Video Lite Distilled
Kandinsky Lab · Kandinsky 6.0 Video
Kandinsky 6.0 Video Lite Distilled generates five-second video with synchronized sound.
Launch date not confirmedDownloadable - Kandinsky 6.0 Video Lite Pretrain
Kandinsky Lab · Kandinsky 6.0 Video
Kandinsky 6.0 Video Lite Pretrain generates five-second video with synchronized sound.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 6.0 Video Pro
Kandinsky Lab · Kandinsky 6.0 Video
Kandinsky 6.0 Video Pro generates five-second video with synchronized sound.
Launch date not confirmedRestricted accessDownloadable - Kandinsky 6.0 Video Pro Distilled
Kandinsky Lab · Kandinsky 6.0 Video
Kandinsky 6.0 Video Pro Distilled generates five-second video with synchronized sound.
Launch date not confirmedDownloadable - Kandinsky 6.0 Video Pro Pretrain
Kandinsky Lab · Kandinsky 6.0 Video
Kandinsky 6.0 Video Pro Pretrain generates five-second video with synchronized sound.
Launch date not confirmedResearch releaseDownloadable - Kandinsky 6.0 VSR
Kandinsky Lab · Kandinsky VSR
A flow-matching diffusion model for increasing the resolution of existing video.
Launch date not confirmedDownloadable - Kandinsky 6.0 VSR Distilled
Kandinsky Lab · Kandinsky VSR
A distilled video super-resolution model that increases resolution with two diffusion evaluations.
Launch date not confirmedDownloadable - Kandinsky Camera Motion Classifier
Kandinsky Lab · Kandinsky video analysis
A VideoMAE-Large classifier that labels camera movements and shot types across a video clip.
Launch date not confirmedResearch releaseDownloadable - Kandinsky S2V 1.0 Lite
Kandinsky Lab · Kandinsky S2V
Kandinsky S2V 1.0 Lite animates a reference image from a supplied speech track.
Launch date not confirmedDownloadable - Kandinsky S2V 1.0 Lite Distilled
Kandinsky Lab · Kandinsky S2V
Kandinsky S2V 1.0 Lite Distilled animates a reference image from a supplied speech track.
Launch date not confirmedDownloadable - Kandinsky S2V 1.0 Pro
Kandinsky Lab · Kandinsky S2V
Kandinsky S2V 1.0 Pro animates a reference image from a supplied speech track.
Launch date not confirmedDownloadable - Kandinsky S2V 1.0 Pro Distilled
Kandinsky Lab · Kandinsky S2V
Kandinsky S2V 1.0 Pro Distilled animates a reference image from a supplied speech track.
Launch date not confirmedDownloadable - Kandinsky Time Adapter
Kandinsky Lab · Kandinsky temporal adapters
A Kandinsky 5 I2V Lite adapter that controls motion speed through frame-rate conditioning and latent-time control.
Launch date not confirmedResearch releaseDownloadable - Kandinsky Video 1.0
Kandinsky Lab · Kandinsky Video
A FusionFrames video model that generates sparse keyframes and interpolates the frames between them.
Launch date not confirmedDownloadable - Kandinsky Video 1.1
Kandinsky Lab · Kandinsky Video
A video model that conditions keyframes on a Kandinsky 3.0 image or supplied image and a text prompt.
Launch date not confirmedDownloadable - Kandinsky Video Motion Predictor
Kandinsky Lab · Kandinsky video analysis
A VideoMAEv2-Base adaptation that predicts three motion-intensity scores for video analysis.
Launch date not confirmedResearch releaseDownloadable - Kandinsky WM 1.0 Autonomous Driving
Kandinsky Lab · Kandinsky WM
A 2B Kandinsky Video Lite world-model adaptation for five-second autonomous-driving scenes.
Launch date not confirmedResearch releaseDownloadable - Kandinsky WM 1.0 General Physics
Kandinsky Lab · Kandinsky WM
A 2B Kandinsky Video Lite world-model adaptation for five-second general-physics scenes.
Launch date not confirmedResearch releaseDownloadable - Kandinsky WM 1.0 Robotics
Kandinsky Lab · Kandinsky WM
A 2B Kandinsky Video Lite world-model adaptation for five-second robotics scenes.
Launch date not confirmedResearch releaseDownloadable - KandiSuperRes 2×
Kandinsky Lab · KandiSuperRes
A dedicated image super-resolution model that increases resolution by a factor of two.
Launch date not confirmedDownloadable - KandiSuperRes 4×
Kandinsky Lab · KandiSuperRes
A dedicated image super-resolution model that increases resolution by a factor of four.
Launch date not confirmedDownloadable - KandiSuperRes Flash
Kandinsky Lab · KandiSuperRes
A distilled 2× super-resolution model combined with the Kandinsky Flash refiner.
Launch date not confirmedDownloadable - KVAE Audio
Kandinsky Lab · KVAE
A continuous audio autoencoder that compresses and reconstructs speech, music and general sound.
Launch date not confirmedResearch releaseDownloadable - KVAE Image 1.0
Kandinsky Lab · KVAE
The original KVAE image autoencoder for compressing RGB images into continuous latents and reconstructing them.
Launch date not confirmedResearch releaseDownloadable - KVAE Image 2.0
Kandinsky Lab · KVAE
The initial KVAE 2.0 image tokenizer retained as a named repository revision.
Launch date not confirmedResearch releaseDownloadable - KVAE Image 2.1
Kandinsky Lab · KVAE
The updated KVAE 2.1 image tokenizer published as the KVAE-2D-2.0 repository's main revision.
Launch date not confirmedResearch releaseDownloadable - KVAE Video 1.0
Kandinsky Lab · KVAE
An earlier causal video autoencoder that converts frame sequences into continuous latent representations.
Launch date not confirmedResearch releaseDownloadable A causal KVAE 2.0 video autoencoder with temporal caching and the t4s16 compression configuration.
Launch date not confirmedResearch releaseDownloadableA causal KVAE 2.0 video autoencoder with temporal caching and the t4s8 compression configuration.
Launch date not confirmedResearch releaseDownloadable- ruDALL-E Emojich
Kandinsky Lab · ruDALL-E
A Malevich adaptation trained on 2,749 text-emoji pairs for emoji-style image generation.
Launch date not confirmedDownloadable - ruDALL-E Kandinsky XXL
Kandinsky Lab · ruDALL-E
A 12B Russian text-to-image model documented in the original ruDALL-E release.
Launch date not confirmedAccess not confirmed - ruDALL-E Malevich XL
Kandinsky Lab · ruDALL-E
A 1.3B Russian text-to-image model trained by Sber AI and SberDevices using the DALL-E approach.
Launch date not confirmedDownloadable