Model reference · open weights
Kandinsky-6.0-VSR-distilled2steps-5s is an open-weight language model from kandinskylab. Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers (BF16) weighs 17.5 GB; the smallest configuration that runs it is RTX 4090 24 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers model card, 2026-10-13
What it is
| Released by | kandinskylab |
|---|---|
| Released | 2026-09-15 |
| Parameters | 1.4B |
| VRAM | 17.5 GB for the weights |
What it runs on
How much memory each request adds isn't estimated yet for this architecture. The weights need at least the cards below, plus room for the context.
| Card | Weights alone |
|---|---|
| RTX 3060 12 GB … RTX 4060 Ti 16 GB 2 smaller cards | does not fit |
| RTX 3090 24 GB | fits |
| RTX 4090 24 GB | fits |
| RTX 5090 32 GB | fits |
| L40S 48 GB | fits |
| A100 80 GB | fits |
| H100 80 GB | fits |
| RTX PRO 6000 Blackwell 96 GB | fits |
| DGX Spark (GB10) 128 GB unified | fits |
| H200 141 GB | fits |
| B200 180 GB | fits |
From the model card
We present Kandinsky 6.0 Video, a family of foundation diffusion models for synchronized text-to-audio-video generation, comprising Kandinsky 6.0 Video Lite (3B parameters) and Kandinsky 6.0 Video Pro (29B parameters). Both models generate 5-second video clips with synchronized 44 kHz audio, including lip-sync, in text-to-audio-video (T2AV) and image-to-audio-video (TI2AV) modes; a plug-in super-resolution model raises the output resolution to Full-HD (1920×1080).
The complete Kandinsky 6.0 video super-resolution pipeline as a single 🧨 Diffusers bundle: the 1.4B SR DiT distilled with π-Flow DX down to 2 model evaluations per tile, the KVAE video VAE and the x2 / x4 latent-upscaler bank. It upscales a video by x2, x2.25 or x4 with tiled diffusion in KVAE latent space. The Diffusers pipeline is video-only; mux the source audio back in if you need it in the output.
This repository contains all components needed for inference with from_pretrained. For the flow-matching teacher with 4 Euler steps per tile, see the flow-matching Diffusers bundle.
import torch
from diffusers import Kandinsky6SRPipeline
from diffusers.utils import export_to_video, load_video
torch._inductor.config.max_autotune = True # required: lets inductor pick flex-attention tiles that fit the SR block mask
pipe = Kandinsky6SRPipeline.from_pretrained("kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()
pipe.transformer.set_attention_backend("flex")
pipe.transformer.compile_repeated_blocks(fullgraph=True)
video = load_video("lq.mp4") # -> list[PIL.Image.Image]
output = pipe(
video=video,
resolution_scale=2.25, # 2, 2.25 or 4
num_inference_steps=2, # must be set explicitly -- the pipeline's own default (4) is for the flow-matching checkpoint
generator=torch.Generator("cuda").manual_seed(1137),
)
export_to_video(output.frames[0], "lq_sr.mp4", fps=24)
| Component | Repository |
|---|---|
| Diffusers bundle, π-Flow distilled (2 evaluations per tile) (this repository) | Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers |
| Diffusers bundle, flow matching | Kandinsky-6.0-VSR-5s-Diffusers |
| Folder | Class | Parameters | Contents |
|---|---|---|---|
transformer/ | Kandinsky6SRTransformer3DModel | 1.41B | π-Flow DX distilled SR DiT (2 evaluations per tile) |
vae/ | Kandinsky6SRVAE | 1.74B | Video VAE (included in this bundle) |
latent_upscaler/ | Kandinsky6SRLatentUpscalerBank (x2 + x4) | 3.65B | x2 and x4 latent upscalers (included in this bundle) |
scheduler/ | PiflowScheduler | — | π-Flow DX policy rollout: nfe = 2, shift = 3.5, 10 grid points, 128 policy substeps |
model_index.json names the pipeline class Kandinsky6SRPipeline.
@misc{kandinsky2026kandinsky60videofoundation,
title={Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation},
author={Team Kandinsky and Julia Agafonova and Bulat Akhmatov and Mikhail Aksyutin and Grigorii Alekseenko and Anastasia Aliaskina and Olga Androsova and Vladimir Arkhipkin and Anna Averchenkova and Alexander Belykh and Serafima Bocharova and Sofiya Bogakovskaya and Anton Bukashkin and Mark Bulygin and Kirill Buzygin and Irina Cheremnykh and Kirill Chernyshev and Mikhail Chernyshov and Vladimir Chernyy and David Chikovani and Georgy Daniltsev and Denis Dimitrov and Anna Dmitrienko and Vladimir Dokholyan and Sergey Emelyanov and Dmitry Ermilov and Georgii Fedorov and Polina Gavrilova and Nikolai Gerasimenko and Aleksandr Gordeev and Andrey Inozemtsev and Andrei Ivaniuta and Alexander Ivanov and Mikhail Karaev and Anastasiia Kargapoltseva and Ivan Kirillov and Nikita Kiselev and Valeria Kobenko and Yury Kolabushin and Denis Koposov and Anatoly Korobov and Vladimir Korviakov and Kirill Kozlov and Denis Krzhivokolskiy and Konstantin Kuklev and Alexander Kunitsyn and Sergey Kuzin and Vladislav Lakhtionov and Alexey Letunovskiy and Maxim Litvinov and Alexander Lyulkov and Georgy Makarov and Kirill Malakhov and Egor Malykh and Mikhail Mamaev and Dmitrii Mikhailov and Polina Mikhailova and Ivan Mikheev and Elizaveta Muromtseva and Nikolai Nazarkin and Tatiana Nikulina and Lev Novitskiy and Stanislav Onuchin and Nikita Osterov and Denis Parkhomenko and Anatoliy Parpara and Vladimir Polovnikov and Konstantin Reznikov and Azat Saginbaev and Nikita Samsonov and Alexander Sentsov and Nikita Shaimov and Artem Sherstyuk and Andrey Shutkin and Egor Silvestrov and Bulat Suleimanov and Matvey Suprunov and Sergey Taranov and Irina Tolstykh and Tatiana Trofimuk and Ilya Trushkin and Aleksandra Tsybina and Olga Varlashina and Viacheslav Vasilev and Ilya Vasiliev and Eugeny Vilisov and Sergey Yakubson and Konstantin Zakharov},
year={2026},
eprint={2610.05608},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.05608},
}
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.