๐ค๐ค๐ค Videotuna is a useful codebase for text-to-video applications.
๐ VideoTuna is the first repo that integrates multiple AI video generation models for text-to-video, image-to-video, text-to-image generation (to the best of our knowledge).
๐ VideoTuna is the first repo that provides comprehensive pipelines in video generation, including pre-training, continuous training, post-training (alignment), and fine-tuning (to the best of our knowledge).
๐ The models of VideoTuna include both U-Net and DiT architectures for visual generation tasks.
๐ A new 3D video VAE, and a controllable facial video generation model will be released soon.
๐ All-in-one framework: Inference and fine-tune up-to-date video generation models.
๐ Pre-training: Build your own foundational text-to-video model.
๐ Continuous training: Keep improving your model with new data.
๐ Domain-specific fine-tuning: Adapt models to your specific scenario.
๐ Concept-specific fine-tuning: Teach your models with unique concepts.
๐ Enhanced language understanding: Improve model comprehension through continuous training.
๐ Post-processing: Enhance the videos with video-to-video enhancement model.
๐ Post-training/Human preference alignment: Post-training with RLHF for more attractive results.
- [2024-11-01] We make the VideoTuna V0.1.0 public!
The 3D video VAE from VideoTuna can accurately compress and reconstruct the input videos with fine details.
Ground Truth | Reconstruction |
Ground Truth | Reconstruction |
Ground Truth | Reconstruction |
Ground Truth | Reconstruction |
Ground Truth | Reconstruction |
Ground Truth | Reconstruction |
Ground Truth | Reconstruction |
Input 1 | Input 2 | Input 3 |
Emotion: Anger | Emotion: Disgust | Emotion: Fear |
Emotion: Happy | Emotion: Sad | Emotion: Surprise |
Emotion: Anger | Emotion: Disgust | Emotion: Fear |
Emotion: Happy | Emotion: Sad | Emotion: Surprise |
Emotion: Anger | Emotion: Disgust | Emotion: Fear |
Emotion: Happy | Emotion: Sad | Emotion: Surprise |
- More demo and applications
- More functionalities such as control modules. (Suggestions are welcome!)
VideoTuna/
โโโ assets # put images for readme
โโโ checkpoints # put model checkpoints here
โโโ configs # model and experimental configs
โโโ data # data processing scripts and dataset files
โโโ docs # documentations
โโโ eval # evaluation scripts
โโโ inputs # input examples for testing
โโโ scripts # train and inference python scripts
โโโ shsripts # train and inference shell scripts
โโโ src # model-related source code
โโโ tests # testing scripts
โโโ tools # some tool scripts
T2V-Models | HxWxL | Checkpoints |
---|---|---|
Mochi | 848x480, 3s | Hugging Face |
CogVideoX-2B | 720x480, 6s | Hugging Face |
CogVideoX-5B | 720x480, 6s | Hugging Face |
Open-Sora 1.0 | 512ร512x16 | Hugging Face |
Open-Sora 1.0 | 256ร256x16 | Hugging Face |
Open-Sora 1.0 | 256ร256x16 | Hugging Face |
VideoCrafter2 | 320x512x16 | Hugging Face |
VideoCrafter1 | 576x1024x16 | Hugging Face |
VideoCrafter1 | 320x512x16 | Hugging Face |
I2V-Models | HxWxL | Checkpoints |
---|---|---|
CogVideoX-5B-I2V | 720x480, 6s | Hugging Face |
DynamiCrafter | 576x1024x16 | Hugging Face |
VideoCrafter1 | 320x512x16 | Hugging Face |
- Note: H: height; W: width; L: length
Please check docs/CHECKPOINTS.md to download all the model checkpoints.
conda create --name videotuna python=3.10 -y
conda activate videotuna
pip install -r requirements.txt
git clone https://github.com/JingyeChen/SwissArmyTransformer
pip install -e SwissArmyTransformer/
git clone https://github.com/tgxs002/HPSv2.git
cd ./HPSv2
pip install -e .
cd ..
conda config --add channels conda-forge
conda install ffmpeg
Please follow docs/CHECKPOINTS.md to download model checkpoints.
After downloading, the model checkpoints should be placed as Checkpoint Structure.
- Inference a set of text-to-video models in one command:
bash tools/video_comparison/compare.sh
- The default mode is to run all models, e.g.,
inference_methods="videocrafter2;dynamicrafter;cogvideoโt2v;cogvideoโi2v;opensora"
- If the users want to inference specific models, modify the
inference_methods
variable incompare.sh
, and list the desired models separated by semicolons. - Also specify the input directory via the
input_dir
variable. This directory should contain aprompts.txt
file, where each line corresponds to a prompt for the video generation. The defaultinput_dir
isinputs/t2v
- The default mode is to run all models, e.g.,
- Inference a set of image-to-video models in one command:
bash tools/video_comparison/compare_i2v.sh
- Inference a specific model, run the corresponding commands as follows:
Task | Model | Command | Length (#frames) | Resolution | Inference Time (s) | GPU Memory (GiB) |
---|---|---|---|---|---|---|
T2V | Mochi | bash shscripts/inference_mochi.sh |
84 | 480x848 | 109.0 | 26 |
I2V | CogVideoX-5b-I2V | bash shscripts/inference_cogVideo_i2v_diffusers.sh |
49 | 576x1024 | 310.4 | 4.78 |
T2V | CogVideoX-2b | bash shscripts/inference_cogVideo_t2v_diffusers.sh |
49 | 576x1024 | 107.6 | 2.32 |
T2V | Open Sora V1.0 | bash shscripts/inference_opensora_v10_16x256x256.sh |
16 | 256x256 | 11.2 | 23.99 |
T2V | VideoCrafter-V2-320x512 | bash shscripts/inference_vc2_t2v_320x512.sh |
16 | 320x512 | 26.4 | 10.03 |
T2V | VideoCrafter-V1-576x1024 | bash shscripts/inference_vc1_t2v_576x1024.sh |
16 | 576x1024 | 91.4 | 14.57 |
I2V | DynamiCrafter | bash shscripts/inference_dc_i2v_576x1024.sh |
16 | 576x1024 | 101.7 | 52.23 |
I2V | VideoCrafter-V1 | bash shscripts/inference_vc1_i2v_320x512.sh |
16 | 320x512 | 26.4 | 10.03 |
T2I | Flux-dev | bash shscripts/inference_flux.sh |
1 | 768x1360 | 238.1 | 1.18 |
T2I | Flux-schnell | bash shscripts/inference_flux.sh |
1 | 768x1360 | 5.4 | 1.20 |
Please follow the docs/datasets.md to try provided toydataset or build your own datasets.
Before started, we assume you have finished the following two preliminary steps:
- Install the environment
- Prepare the dataset
- Download the checkpoints and get these two checkpoints
ll checkpoints/videocrafter/t2v_v2_512/model.ckpt
ll checkpoints/stablediffusion/v2-1_512-ema/model.ckpt
First, run this command to convert the VC2 checkpoint as we make minor modifications on the keys of the state dict of the checkpoint. The converted checkpoint will be automatically save at checkpoints/videocrafter/t2v_v2_512/model_converted.ckpt
.
python tools/convert_checkpoint.py --input_path checkpoints/videocrafter/t2v_v2_512/model.ckpt
Second, run this command to start training on the single GPU. The training results will be automatically saved at results/train/${CURRENT_TIME}_${EXPNAME}
bash shscripts/train_videocrafter_v2.sh
We support lora finetuning to make the model to learn new concepts/characters/styles.
- Example config file:
configs/001_videocrafter2/vc2_t2v_lora.yaml
- Training lora based on VideoCrafter2:
bash shscripts/train_videocrafter_lora.sh
- Inference the trained models:
bash shscripts/inference_vc2_t2v_320x512_lora.sh
We support open-sora finetuning, you can simply run the following commands:
# finetune the Open-Sora v1.0
bash shscripts/train_opensorav10.sh
We support VBench evaluation to evaluate the T2V generation performance. Please check eval/README.md for details.
We thank the following repos for sharing their awesome models and codes!
- Mochi: A new SOTA in open-source video generation models
- VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models
- VideoCrafter1: Open Diffusion Models for High-Quality Video Generation
- DynamiCrafter: Animating Open-domain Images with Video Diffusion Priors
- Open-Sora: Democratizing Efficient Video Production for All
- CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer
- VADER: Video Diffusion Alignment via Reward Gradients
- VBench: Comprehensive Benchmark Suite for Video Generative Models
- Flux: Text-to-image models from Black Forest Labs.
- SimpleTuner: A fine-tuning kit for text-to-image generation.
- LLMs-Meet-MM-Generation: A paper collection of utilizing LLMs for multimodal generation (image, video, 3D and audio).
- MMTrail: A multimodal trailer video dataset with language and music descriptions.
- Seeing-and-Hearing: A versatile framework for Joint VA generation, V2A, A2V, and I2A.
- Self-Cascade: A Self-Cascade model for higher-resolution image and video generation.
- ScaleCrafter and HiPrompt: Free method for higher-resolution image and video generation.
- FreeTraj and FreeNoise: Free method for video trajectory control and longer-video generation.
- Follow-Your-Emoji, Follow-Your-Click, and Follow-Your-Pose: Follow family for controllable video generation.
- Animate-A-Story: A framework for storytelling video generation.
- LVDM: Latent Video Diffusion Model for long video generation and text-to-video generation.
Please follow CC-BY-NC-ND. If you want a license authorization, please contact the project leads Yingqing He ([email protected]) and Yazhou Xing ([email protected]).
@software{videotuna,
author = {Yingqing He and Yazhou Xing and Zhefan Rao and Haoyu Wu and Zhaoyang Liu and Jingye Chen and Pengjun Fang and Jiajun Li and Liya Ji and Runtao Liu and Xiaowei Chi and Yang Fei and Guocheng Shao and Yue Ma and Qifeng Chen},
title = {VideoTuna: A Powerful Toolkit for Video Generation with Model Fine-Tuning and Post-Training},
month = {Nov},
year = {2024},
url = {https://github.com/VideoVerses/VideoTuna}
}