Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers

<div align="center"> <picture> <img src="assets/KANDINSKY_LOGO_1_BLACK.png"> </picture> </div> <div align="center"> <a href="https://habr.com/ru/companies/sberbank/articles/951800/">Habr</a> | <a href="https://kandinskylab.ai/">Project Page</a> | <a href="https://arxiv.org/abs/2511.14993">Technical Report</a> | <a href="https://github.com/kandinskylab/Kandinsky-5">Original Github</a> | <a href="https://huggingface.co/collections/kandinskylab/kandinsky-50-video-lite-diffusers"> 🤗 Diffusers</a> </div>

---

<h1>Kandinsky 5.0 T2V Lite - Diffusers</h1>

This repository provides the 🤗 Diffusers integration for Kandinsky 5.0 T2V Lite - a lightweight video generation model (2B parameters) that ranks #1 among open-source models in its class.

Project Updates

• 🔥 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.
• 🚀 Diffusers Integration: Now available with easy-to-use 🤗 Diffusers pipeline!

Kandinsky 5.0 T2V 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

Basic Usage

import torch
from diffusers import Kandinsky5T2VPipeline
from diffusers.utils import export_to_video

# Load the pipeline
pipe = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers", 
    torch_dtype=torch.bfloat16
)
pipe = pipe.to("cuda")

# Generate video
prompt = "A cat and a dog baking a cake together in a kitchen."
negative_prompt = "Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards"

output = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    height=512,
    width=768,
    num_frames=121,
    num_inference_steps=50,
    guidance_scale=5.0,
).frames[0]

## Save the video
export_to_video(output, "output.mp4", fps=24, quality=9)

Using Different Model Variants

import torch
from diffusers import Kandinsky5T2VPipeline

# 5s SFT model (highest quality)
pipe_sft = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers",
    torch_dtype=torch.bfloat16
)

# 5s Distilled 16-step model (fastest)
pipe_distill = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers", 
    torch_dtype=torch.bfloat16
)

# 5s No-CFG model (balanced speed/quality)
pipe_nocfg = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-nocfg-5s-Diffusers",
    torch_dtype=torch.bfloat16
)

# 5s Pretrain model (most diverse)
pipe_pretrain = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-pretrain-5s-Diffusers",
    torch_dtype=torch.bfloat16
)

# 10s SFT model (highest quality)
pipe_sft = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-sft-10s-Diffusers",
    torch_dtype=torch.bfloat16
)

# 10s Distilled 16-step model (fastest)
pipe_distill = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-10s-Diffusers", 
    torch_dtype=torch.bfloat16
)

# 10s No-CFG model (balanced speed/quality)
pipe_nocfg = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-nocfg-10s-Diffusers",
    torch_dtype=torch.bfloat16
)

# 10s Pretrain model (most diverse)
pipe_pretrain = Kandinsky5T2VPipeline.from_pretrained(
    "kandinskylab/Kandinsky-5.0-T2V-Lite-pretrain-10s-Diffusers",
    torch_dtype=torch.bfloat16
)

Architecture

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

<div align="center"> <img width="1600" height="477" alt="Pipeline Architecture" src="https://github.com/user-attachments/assets/17fc2eb5-05e3-4591-9ec6-0f6e1ca397b3" /> </div> <div align="center"> <img width="800" height="406" alt="Model Architecture" src="https://github.com/user-attachments/assets/f3006742-e261-4c39-b7dc-e39330be9a09" /> </div>

Examples

Kandinsky 5.0 T2V Lite SFT

<table border="0" style="width: 200; text-align: left; margin-top: 20px;"> <tr> <td> <video src="https://github.com/user-attachments/assets/bc38821b-f9f1-46db-885f-1f70464669eb" width=200 controls autoplay loop></video> </td> <td> <video src="https://github.com/user-attachments/assets/9f64c940-4df8-4c51-bd81-a05de8e70fc3" width=200 controls autoplay loop></video> </td> <tr> <td> <video src="https://github.com/user-attachments/assets/77dd417f-e0bf-42bd-8d80-daffcd054add" width=200 controls autoplay loop></video> </td> <td> <video src="https://github.com/user-attachments/assets/385a0076-f01c-4663-aa46-6ce50352b9ed" width=200 controls autoplay loop></video> </td> <tr> <td> <video src="https://github.com/user-attachments/assets/7c1bcb31-cc7d-4385-9a33-2b0cc28393dd" width=200 controls autoplay loop></video> </td> <td> <video src="https://github.com/user-attachments/assets/990a8a0b-2df1-4bbc-b2e3-2859b6f1eea6" width=200 controls autoplay loop></video> </td> </tr> </table> Kandinsky 5.0 T2V Lite Distill <table border="0" style="width: 200; text-align: left; margin-top: 20px;"> <tr> <td> <video src="https://github.com/user-attachments/assets/861342f9-f576-4083-8a3b-94570a970d58" width=200 controls autoplay loop></video> </td> <td> <video src="https://github.com/user-attachments/assets/302e4e7d-781d-4a58-9b10-8c473d469c4b" width=200 controls autoplay loop></video> </td> <tr> <td> <video src="https://github.com/user-attachments/assets/3e70175c-40e5-4aec-b506-38006fe91a76" width=200 controls autoplay loop></video> </td> <td> <video src="https://github.com/user-attachments/assets/b7da85f7-8b62-4d46-9460-7f0e505de810" width=200 controls autoplay loop></video> </td> </table> Results Side-by-Side Evaluation The evaluation is based on the expanded prompts from the Movie Gen benchmark. <table border="0" style="width: 400; text-align: left; margin-top: 20px;"> <tr> <td> <img src="assets/sbs/kandinsky_5_video_lite_vs_sora.jpg" width=400 ></img> </td> <td> <img src="assets/sbs/kandinsky_5_video_lite_vs_wan_2.1_14B.jpg" width=400 ></img> </td> <tr> <td> <img src="assets/sbs/kandinsky_5_video_lite_vs_wan_2.2_5B.jpg" width=400 ></img> </td> <td> <img src="assets/sbs/kandinsky_5_video_lite_vs_wan_2.2_A14B.jpg" width=400 ></img> </td> <tr> <td> <img src="assets/sbs/kandinsky_5_video_lite_vs_wan_2.1_1.3B.jpg" width=400 ></img> </td> </table> Distill Side-by-Side Evaluation <table border="0" style="width: 400; text-align: left; margin-top: 20px;"> <tr> <td> <img src="assets/sbs/kandinsky_5_video_lite_5s_vs_kandinsky_5_video_lite_distill_5s.jpg" width=400 ></img> </td> <td> <img src="assets/sbs/kandinsky_5_video_lite_10s_vs_kandinsky_5_video_lite_distill_10s.jpg" width=400 ></img> </td> </table> VBench Results <div align="center"> <picture> <img src="assets/vbench.png"> </picture> </div> Beta Testing You can apply to participate in the beta testing of the Kandinsky Video Lite via the telegram bot.

@misc{arkhipkin2026kandinsky50familyfoundation,
      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={2026},
      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}, 
}