Kandinsky 4.0 T2V Flash is a text-to-video generation model based on latent diffusion for 480p resolution, that can generate 12 second videos in 480p resolution in 11 seconds on a single NVIDIA H100 gpu. The pipeline consist of 3D causal CogVideoX VAE, text embedder T5-V1.1-XXL and our trained MMDiT-like transformer model.
A serious problem for all diffusion models, and especially video generation models, is the generation speed. To solve this problem, we used the Latent Adversarial Diffusion Distillation (LADD) approach, proposed for distilling image generation models and first described in the article from Stability AI and tested by us when training the Kandinsky 3.1 image generation model. The distillation pipeline itself involves additional training of the diffusion model in the GAN pipeline, i.e. joint training of the diffusion generator with the discriminator.
Architecture
For training Kandinsky 4.0 T2V Flash we used the following architecture of diffusion transformer, based on MMDiT proposed in Stable Diffusion 3.
For training flash version we used the following architecture of discriminator. Discriminator head structure resembles half of an MMDiT block.
1import torch
2from IPython.display import Video
3from kandinsky import get_T2V_pipeline
45device_map ={6"dit": torch.device('cuda:0'),7"vae": torch.device('cuda:0'),8"text_embedder": torch.device('cuda:0')9}1011pipe = get_T2V_pipeline(device_map)1213images = pipe(14 seed=42,15 time_length=12,16 width =672,17 height =384,18 save_path="./test.mp4",19 text="Several giant wooly mammoths approach treading through a snowy meadow, their long wooly fur lightly blows in the wind as they walk, snow covered trees and dramatic snow capped mountains in the distance",20)2122Video("./test.mp4")
Examples of usage and more detailed parameters description are in the examples.ipynb notebook.
Make sure that you have weights folder with weights of all models.
python -m torch.distributed.launch --nnodes n --nproc-per-node m run_inference_distil.py
where n is a number of nodes you have and m is a number of gpus on these nodes. The code was tested with n=1 and m=8, so this is preferable parameters.
In distributed setting the DiT models are parallelized using tensor parallel on all gpus, which enables a significant speedup.
To run this examples from terminal without tensor parallel:
python run_inference_distil.py
Authors
Project Leader
Denis Dimitrov
Scientific Consultants
Andrey Kuznetsov, Sergey Markov
Training Pipeline & Model Pretrain & Model Distillation
Vladimir Arkhipkin, Novitskiy Lev, Maria Kovaleva
Model Architecture
Vladimir Arkhipkin, Maria Kovaleva, Zein Shaheen, Arsen Kuzhamuratov, Nikolay Gerasimenko, Mikhail Zhirnov, Alexandr Gambashidze, Konstantin Sobolev
Data Pipeline
Ivan Kirillov, Andrei Shutkin, Kirill Chernishev, Julia Agafonova, Denis Parkhomenko
Video-to-audio model
Zein Shaheen, Arseniy Shakhmatov, Denis Parkhomenko
Quality Assessment
Nikolay Gerasimenko, Anna Averchenkova, Victor Panshin, Vladislav Veselov, Pavel Perminov, Vladislav Rodionov, Sergey Skachkov, Stepan Ponomarev
Other Contributors
Viacheslav Vasilev, Andrei Filatov, Gregory Leleytner
