kandinsky-5

In this repository, we provide a family of diffusion models to generate a video or an image given a textual prompt and/or image.

https://github.com/user-attachments/assets/b06f56de-1b05-4def-a611-1a3159ed71b0

Project Updates

• 🔥 2025/12/12: Kandinsky 5.0 Video Pro is ranked as Top-1 open-source Text-to-Video model at LMArena.
• 🔥 2025/12/03: our Kandinsky 5.0 Video Pro and Image Lite models are accepted to diffusers
• 🔥 2025/11/26: our simplified LoRa training is available on kandinsky-5-lora-train
• 🔥 2025/11/24: LoRas for Camera control is open-sourced: Lite LoRAs and Pro LoRAs. Inference code is avaibale in examples/inference_examples_i2v_lora.ipynb and examples/inference_examples_t2v_lora.ipynb
• 🔥 2025/11/20: Kandinsky 5.0 Video Pro is open-sourced. T2V & I2V models are available.
• 🔥 2025/11/15: Kandinsky 5.0 Lite I2V & Kandinsky 5.0 Lite T2I models are open-sourced.
• 🔥 2025/10/19: Further VAE tiling optimization. NF4 version of Qwen2.5-VL from Bitsandbytes is supported. Flash Attention 2, Flash Attention 2, Sage Attention or SDPA can be selected for 5-seconds generation using option --attention_engine. Now generation should work on the GPUS with 12 GB of memory. Kandinsky 5 Video Lite is accepted to diffusers.
• 🔥 2025/10/7: The ComfyUI README file has been updated. SDPA support has been added, allowing you to run our code without Flash attention. Magcache support for nocfg checkpoints has been added, allowing Magcache support for sft and nocfg checkpoints. Memory consumption in the VAE has been reduced, with the entire pipeline now running at 24 GB with offloading.
• 🔥 2025/09/29: We have open-sourced Kandinsky 5.0 T2V Lite a lite (2B parameters) version of Kandinsky 5.0 Video text-to-video generation model. Released checkpoints: kandinsky5lite_t2v_pretrain_5s, kandinsky5lite_t2v_pretrain_10s, kandinsky5lite_t2v_sft_5s, kandinsky5lite_t2v_sft_10s, kandinsky5lite_t2v_nocfg_5s, kandinsky5lite_t2v_nocfg_10s, kandinsky5lite_t2v_distilled16steps_5s, kandinsky5lite_t2v_distilled16steps_10s contains weight from pretrain, supervised finetuning, cfg distillation and diffusion distillation into 16 steps. 5s checkpoints are capable of generating videos up to 5 seconds long. 10s checkpoints is faster models checkpoints trained with NABLA algorithm and capable to generate videos up to 10 seconds long.

Community Works

If your research or project builds upon Kandinsky 5, and you would like more people to see it, please inform us.

• bghira/SimpleTuner now supports Kandinsky 5.0 LoRA and even full-rank training.
• CacheDiT offers Fully Cache Acceleration support for Kandinsky-5 with DBCache, TaylorSeer and Cache CFG. Visit their example for more details.

Table of Contents

1. Kandinsky 5.0 Video Pro
2. Kandinsky 5.0 Video Lite
3. Kandinsky 5.0 Image Lite
4. Kandinsky 5.0 Image Editing
5. Quickstart & Run examples

Kandinsky 5.0 Video Pro

Kandinsky 5.0 Video Pro is a line-up of 19B models that generates high-quality HD videos from English and Russian prompts with controllable camera motion.

We provide several Text-to-Video model variants, each optimized for different use cases:

• SFT model — delivers the highest generation quality;

• Pretrain model — designed for fine-tuning by researchers and enthusiasts.

All models are available in two versions: for generating 5-second and 10-second videos.

Additionally, we provide Image-to-Video model capable to generate video given input image and text prompt.

Pipeline

Latent diffusion pipeline with Flow Matching.

Diffusion Transformer (DiT) as the main generative backbone with cross-attention to text embeddings.

• Qwen2.5-VL and CLIP provides text embeddings.

• HunyuanVideo 3D VAE encodes/decodes video into a latent space.

• DiT is the main generative module using cross-attention to condition on text.

Model Zoo

Model	config	video duration	NFE	Checkpoint	Latency*
Kandinsky 5.0 T2V Pro SFT 5s HD	configs/k5_pro_t2v_5s_sft_hd.yaml	5s	100	🤗 HF	1241
Kandinsky 5.0 T2V Pro SFT 10s HD	configs/k5_pro_t2v_10s_sft_hd.yaml	10s	100	🤗 HF	-
Kandinsky 5.0 T2V Pro SFT 5s SD	configs/k5_pro_t2v_5s_sft_sd.yaml	5s	100	🤗 HF	560
Kandinsky 5.0 T2V Pro SFT 10s SD	configs/k5_pro_t2v_10s_sft_sd.yaml	10s	100	🤗 HF	1158
Kandinsky 5.0 T2V Pro pretrain 5s HD	-	5s	100	🤗 HF	1241
Kandinsky 5.0 T2V Pro pretrain 10s HD	-	10s	100	🤗 HF	-
Kandinsky 5.0 T2V Pro pretrain 5s SD	-	5s	100	🤗 HF	560
Kandinsky 5.0 T2V Pro pretrain 10s SD	-	10s	100	🤗 HF	1158
Kandinsky 5.0 I2V Pro HD 5s	configs/k5_pro_i2v_5s_sft_hd.yaml	5s	100	🤗 HF	-
Kandinsky 5.0 I2V Pro SD 5s	configs/k5_pro_i2v_5s_sft_sd.yaml	5s	100	🤗 HF	-

*Latency was measured after the second inference run. The first run of the model can be slower due to the compilation process. Inference was measured on an NVIDIA H100 GPU with 80 GB of memory, using CUDA 12.8.1 and PyTorch 2.8. For 5-second models Flash Attention 3 was used.

Examples:

Results:

Side-by-Side evaluation

Kandinsky 5.0 Video Lite

Kandinsky 5.0 T2V Lite is a lightweight video generation model (2B parameters) that ranks #1 among open-source models in its class. It outperforms larger Wan models (5B and 14B) and offers the best understanding of Russian concepts in the open-source ecosystem.

We provide 8 model variants, each optimized for different use cases:

• SFT model — delivers the highest generation quality;

• CFG-distilled — runs 2× faster;

• Diffusion-distilled — enables low-latency generation with minimal quality loss (6× faster);

• Pretrain model — designed for fine-tuning by researchers and enthusiasts.

All models are available in two versions: for generating 5-second and 10-second videos.

Additionally, we provide Image-to-Video model capable to generate video given input image and text prompt.

Model Zoo

Model	config	video duration	NFE	Checkpoint	Latency*
Kandinsky 5.0 T2V Lite SFT 5s	configs/k5_lite_t2v_5s_sft_sd.yaml	5s	100	🤗 HF	139 s
Kandinsky 5.0 T2V Lite SFT 10s	configs/k5_lite_t2v_10s_sft_sd.yaml	10s	100	🤗 HF	224 s
Kandinsky 5.0 T2V Lite pretrain 5s	configs/k5_lite_t2v_5s_pretrain_sd.yaml	5s	100	🤗 HF	139 s
Kandinsky 5.0 T2V Lite pretrain 10s	configs/k5_lite_t2v_10s_pretrain_sd.yaml	10s	100	🤗 HF	224 s
Kandinsky 5.0 T2V Lite no-CFG 5s	configs/k5_lite_t2v_5s_nocfg_sd.yaml	5s	50	🤗 HF	77 s
Kandinsky 5.0 T2V Lite no-CFG 10s	configs/k5_lite_t2v_10s_nocfg_sd.yaml	10s	50	🤗 HF	124 s
Kandinsky 5.0 T2V Lite distill 5s	configs/k5_lite_t2v_5s_distil_sd.yaml	5s	16	🤗 HF	35 s
Kandinsky 5.0 T2V Lite distill 10s	configs/k5_lite_t2v_10s_distil_sd.yaml	10s	16	🤗 HF	61 s
Kandinsky 5.0 I2V Lite 5s	configs/k5_lite_i2v_5s_sft_sd.yaml	5s	100	🤗 HF	139 s

*Latency was measured after the second inference run. The first run of the model can be slower due to the compilation process. Inference was measured on an NVIDIA H100 GPU with 80 GB of memory, using CUDA 12.8.1 and PyTorch 2.8. For 5-second models Flash Attention 3 was used.

Examples:

Kandinsky 5.0 T2V Lite SFT

Kandinsky 5.0 T2V Lite Distill

Results:

Side-by-Side evaluation

The evaluation is based on the expanded prompts from the Movie Gen benchmark, which are available in the expanded_prompt column of the benchmark/moviegen_bench.csv file.

Distill Side-by-Side evaluation

Kandinsky 5.0 Image Lite

Kandinsky 5.0 Image Lite is a line-up of 6B image generation models with the following capabilities:

• 1K resulution (1280x768, 1024x1024 and others).

• High visual quality

• Strong text-writing

• Russian concepts understanding

Model Zoo

Model	config	NFE	Checkpoint	Latency*
Kandinsky 5.0 T2I Lite	configs/k5_lite_t2i_sft_hd.yaml	100	🤗 HF	13 s
Kandinsky 5.0 T2I Lite pretrain	-	100	🤗 HF	13 s

*Latency was measured after the second inference run. The first run of the model can be slower due to the compilation process. Inference was measured on an NVIDIA H100 GPU with 80 GB of memory, using CUDA 12.8.1 and PyTorch 2.8. For 5-second models Flash Attention 3 was used.

Examples:

Results

Results:

Side-by-Side evaluation

Kandinsky 5.0 Image Editing

Kandinsky 5.0 Image Editing is a line-up of 6B image editing models with the following capabilities:

• 1K resulution (1280x768, 1024x1024 and others).

• High visual quality

• Strong text-writing

• Russian concepts understanding

Model Zoo

Model	config	NFE	Checkpoint	Latency*
Kandinsky 5.0 T2I Editing	configs/k5_lite_i2i_sft_hd.yaml	100	🤗 HF	-
Kandinsky 5.0 T2I Editing pretrain	-	100	🤗 HF	-

*Latency was measured after the second inference run. The first run of the model can be slower due to the compilation process. Inference was measured on an NVIDIA H100 GPU with 80 GB of memory, using CUDA 12.8.1 and PyTorch 2.8. For 5-second models Flash Attention 3 was used.

Examples:

Results:

Side-by-Side evaluation

Quickstart

Installation

Clone the repo:

git clone https://github.com/kandinskylab/kandinsky-5.git
cd kandinsky-5

Install dependencies:

pip install -r requirements.txt

To improve inference performance on NVidia Hopper GPUs, we recommend installing Flash Attention 3.

Model Download

python download_models.py

use models argument to download some specific models, otherwise all models will be downloaded

example to download only kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s and kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s:

python download_models.py --models kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s,kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s

Run Kandinsky 5.0 T2V Lite SFT 5s

python test.py --prompt "A dog in red hat"

Run Kandinsky 5.0 T2V Lite SFT 10s

python test.py --config ./configs/k5_lite_t2v_10s_sft_sd.yaml --prompt "A dog in red hat" --video_duration 10 

Run Kandinsky 5.0 I2V Lite 5s

python test.py --config ./configs/k5_lite_i2v_5s_sft_sd.yaml --prompt "The bear plays balalaika." --image "./assets/test_image.jpg" --video_duration 5

Run Kandinsky 5.0 T2I Lite

python test.py --config ./configs/k5_lite_t2i_sft_hd.yaml --prompt "A dog in a red hat" --width=1280 --height=768

T2V Inference

import torch
from kandinsky import get_T2V_pipeline

device_map = {
    "dit": torch.device('cuda:0'), 
    "vae": torch.device('cuda:0'), 
    "text_embedder": torch.device('cuda:0')
}

pipe = get_T2V_pipeline(device_map, conf_path="configs/k5_lite_t2v_5s_sft_sd.yaml")

images = pipe(
    seed=42,
    time_length=5,
    width=768,
    height=512,
    save_path="./test.mp4",
    text="A cat in a red hat",
)

I2V Inference

import torch
from kandinsky import get_I2V_pipeline

device_map = {
    "dit": torch.device('cuda:0'), 
    "vae": torch.device('cuda:0'), 
    "text_embedder": torch.device('cuda:0')
}

pipe = get_I2V_pipeline(device_map, conf_path="configs/k5_lite_i2v_5s_sft_sd.yaml")

images = pipe(
    seed=42,
    time_length=5,
    save_path='./test.mp4',
    text="The bear plays balalaika.",
    image = "assets/test_image.jpg",
)

T2I Inference

import torch
from kandinsky import get_T2I_pipeline

device_map = {
    "dit": torch.device('cuda:0'), 
    "vae": torch.device('cuda:0'), 
    "text_embedder": torch.device('cuda:0')
}

pipe = get_T2I_pipeline(device_map, conf_path="configs/k5_lite_t2i_sft_hd.yaml")

images = pipe(
    seed=42,
    save_path='./test.png',
    text="A cat in a red hat with a label 'HELLO'"
)

I2I Inference

import torch
from kandinsky import get_I2I_pipeline

device_map = {
    "dit": torch.device('cuda:0'), 
    "vae": torch.device('cuda:0'), 
    "text_embedder": torch.device('cuda:0')
}

pipe = get_I2I_pipeline(
    resolution=1024, offload=True,
    device_map=device_map,
)
out = pipe(
    "Replace the cat with a husky, leave the rest unchanged",
    image='./assets/cat_in_hat.png'
)

Please, refer to examples folder for more examples in various notebooks.

Distributed Inference

For a faster inference, we also provide the capability to perform inference in a distributed way:

NUMBER_OF_NODES=1
NUMBER_OF_DEVICES_PER_NODE=1 / 2 / 4
python -m torch.distributed.launch --nnodes $NUMBER_OF_NODES --nproc-per-node $NUMBER_OF_DEVICES_PER_NODE test.py

Optimized Inference

Offloading

For less memory consumption you can use offloading of the models.

python test.py --prompt "A dog in red hat" --offload

Magcache

Also we provide Magcache inference for faster generations (now available for sft 5s and sft 10s checkpoints).

python test.py --prompt "A dog in red hat" --magcache

Qwen encoder quantization

To reduce GPU memory needed for Qwen encoder we provide option to use NF4-quantized version from bitsandbytes.

python test.py --prompt "A dog in red hat" --qwen_quantization

Attention engine selection

Depending on your hardware you can use the follwing full attention algorithm implementation:

• PyTorch SDPA
• Flash Attention 2
• Flash Attention 3
• Sage Attention

The attention algorithm can be selected using an option "--attention_engine" of test.py script for 5 second (and less) video generation. For 10-second generation we use sparse attention algorithm NABLA.

Note that currently (19 Oct. 2025) version build from source contains a bug and produces noisy output. A temporary workaround to fix it is decribed here.

python test.py --prompt "A dog in red hat" --attention_engine=flash_attention_3

python test.py --prompt "A dog in red hat" --attention_engine=flash_attention_2

python test.py --prompt "A dog in red hat" --attention_engine=sdpa

python test.py --prompt "A dog in red hat" --attention_engine=sage

By default we use option --attention_engine=auto which enables automatic selection of the most optimal algorithm installed in your system.

ComfyUI

See the instruction here

Beta testing

You can apply to participate in the beta testing of the Kandinsky Video Lite via the telegram bot.

📑 Todo List

• Kandinsky 5.0 Video Pro
  • Checkpoints
    • sft
    • pretrain
    • rl
    • distil 16 steps
    • I2V
  • ComfyUI integration
  • Diffusers integration
  • Caching acceleration support
  • Multi-GPU Inference code of the models
• Kandinsky 5.0 Video Lite
  • Checkpoints
    • sft
    • pretrain
    • rl
    • cfg distil
    • distil 16 steps
    • autoregressive generation
    • I2V
  • ComfyUI integration
  • Diffusers integration
  • Caching acceleration support
  • Multi-GPU Inference code of the models
• Kandinsky 5.0 Image Lite
  • Checkpoints
    • rl
    • pretrain
  • ComfyUI integration
  • Diffusers integration
  • Caching acceleration support
  • Multi-GPU Inference code of the models
• Kandinsky 5.0 Image Editing
  • Checkpoints
    • sft
    • pretrain
  • ComfyUI integration
  • Diffusers integration
  • Multi-GPU Inference code of the models
• Technical report

Authors

Core Contributors:

• Video: Alexey Letunovskiy, Maria Kovaleva, Lev Novitskiy, Denis Koposov, Dmitrii Mikhailov, Anastasiia Kargapoltseva, Anna Dmitrienko, Anastasia Maltseva
• Image & Editing: Nikolai Vaulin, Nikita Kiselev, Alexander Varlamov
• Pre-training Data: Ivan Kirillov, Andrey Shutkin, Nikolai Vaulin, Ilya Vasiliev
• Post-training Data: Julia Agafonova, Anna Averchenkova, Olga Kim
• Research Consolidation & Paper: Viacheslav Vasilev, Vladimir Polovnikov

Contributors: Yury Kolabushin, Kirill Chernyshev, Alexander Belykh, Mikhail Mamaev, Anastasia Aliaskina, Kormilitsyn Semen, Tatiana Nikulina, Olga Vdovchenko, Polina Mikhailova, Polina Gavrilova, Nikita Osterov, Bulat Akhmatov

Track Leaders: Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko

Project Supervisor: Denis Dimitrov

Citation

@misc{arkhipkin2025kandinsky50familyfoundation,
      title={Kandinsky 5.0: A Family of Foundation Models for Image and Video Generation}, 
      author={Vladimir Arkhipkin and Vladimir Korviakov and Nikolai Gerasimenko and Denis Parkhomenko and Viacheslav Vasilev and Alexey Letunovskiy and Nikolai Vaulin and Maria Kovaleva and Ivan Kirillov and Lev Novitskiy and Denis Koposov and Nikita Kiselev and Alexander Varlamov and Dmitrii Mikhailov and Vladimir Polovnikov and Andrey Shutkin and Julia Agafonova and Ilya Vasiliev and Anastasiia Kargapoltseva and Anna Dmitrienko and Anastasia Maltseva and Anna Averchenkova and Olga Kim and Tatiana Nikulina and Denis Dimitrov},
      year={2025},
      eprint={2511.14993},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.14993}, 
}

@misc{mikhailov2025nablanablaneighborhoodadaptiveblocklevel,
      title={$\nabla$NABLA: Neighborhood Adaptive Block-Level Attention}, 
      author={Dmitrii Mikhailov and Aleksey Letunovskiy and Maria Kovaleva and Vladimir Arkhipkin
              and Vladimir Korviakov and Vladimir Polovnikov and Viacheslav Vasilev
              and Evelina Sidorova and Denis Dimitrov},
      year={2025},
      eprint={2507.13546},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.13546}, 
}

Acknowledgements

We gratefully acknowledge the open-source projects and research that made Kandinsky 5.0 possible:

• PyTorch — for model training and inference.
• FlashAttention 3 — for efficient attention and faster inference.
• Qwen2.5-VL — for providing high-quality text embeddings.
• CLIP — for robust text–image alignment.
• HunyuanVideo — for video latent encoding and decoding.
• MagCache — for accelerated inference.
• ComfyUI — for integration into node-based workflows.

We deeply appreciate the contributions of these communities and researchers to the open-source ecosystem.