Awesome

Text-To-Video-Finetuning

Finetune ModelScope's Text To Video model using Diffusers 🧨

Important Update 2023-12-14

First of all a note from me. Thank you guys for your support, feedback, and journey through discovering the nascent, innate potential of video Diffusion Models.

@damo-vilab Has released a repository for finetuning all things Video Diffusion Models, and I recommend their implementation over this repository. https://github.com/damo-vilab/i2vgen-xl

https://github.com/ExponentialML/Text-To-Video-Finetuning/assets/59846140/55608f6a-333a-458f-b7d5-94461c5da8bb

This repository will no longer be updated, but will instead be archived for researchers & builders that wish to bootstrap their projects. I will be leaving the issues, pull requests, and all related things for posterity purposes.

Thanks again!

Updates

• 2023-7-12: You can now train a LoRA that is compatibile with the webui extension! See instructions here.
• 2023-4-17: You can now convert your trained models from diffusers to .ckpt format for A111 webui. Thanks @kabachuha!
• 2023-4-8: LoRA Training released! Checkout configs/v2/lora_training_config.yaml for instructions.
• 2023-4-8: Version 2 is released!
• 2023-3-29: Added gradient checkpointing support.
• 2023-3-27: Support for using Scaled Dot Product Attention for Torch 2.0 users.

Getting Started

Requirements & Installation

git clone https://github.com/ExponentialML/Text-To-Video-Finetuning.git
cd Text-To-Video-Finetuning
git lfs install
git clone https://huggingface.co/damo-vilab/text-to-video-ms-1.7b ./models/model_scope_diffusers/

Other Models

Alternatively, you can train starting from other models made by the community.

Create Conda Environment (Optional)

It is recommended to install Anaconda.

Windows Installation: https://docs.anaconda.com/anaconda/install/windows/

Linux Installation: https://docs.anaconda.com/anaconda/install/linux/

conda create -n text2video-finetune python=3.10
conda activate text2video-finetune

Python Requirements

pip install -r requirements.txt

Hardware

All code was tested on Python 3.10.9 & Torch version 1.13.1 & 2.0.

It is highly recommended to install >= Torch 2.0. This way, you don't have to install Xformers or worry about memory performance.

If you don't have Xformers enabled, you can follow the instructions here: https://github.com/facebookresearch/xformers

Recommended to use a RTX 3090, but you should be able to train on GPUs with <= 16GB ram with:

• Validation turned off.
• Xformers or Torch 2.0 Scaled Dot-Product Attention
• Gradient checkpointing enabled.
• Resolution of 256.
• Hybrid LoRA training.
• Training only using LoRA with ranks between 4-16.

Preprocessing your data

Using Captions

You can use caption files when training on images or video. Simply place them into a folder like so:

Images: /images/img.png /images/img.txt Videos: /videos/vid.mp4 | /videos/vid.txt

Then in your config, make sure to have -folder enabled, along with the root directory containing the files.

Process Automatically

You can automatically caption the videos using the Video-BLIP2-Preprocessor Script

Configuration

The configuration uses a YAML config borrowed from Tune-A-Video reposotories.

All configuration details are placed in configs/v2/train_config.yaml. Each parameter has a definition for what it does.

How would you recommend I proceed with making a config with my data?

I highly recommend (I did this myself) going to configs/v2/train_config.yaml. Then make a copy of it and name it whatever you wish my_train.yaml.

Then, follow each line and configure it for your specific use case.

The instructions should be clear enough to get you up and running with your dataset, but feel free to ask any questions in the discussion board.

Training a LoRA

Please read this section carefully if you are training a LoRA model

You can also train a LoRA that is compatible with the webui extension. By default it's set to 'cloneofsimo', which was the first LoRA implementation for Stable Diffusion.

This ('cloneofsimo') version you can use in the inference.py file in this repository. It is not compatible with the webui.

To train and use a LoRA with the webui, change the lora_version to "stable_lora" in your config if you already have one made.

This will train an A1111 webui extension compatibile LoRA. You can get started at configs/v2/stable_lora_config.yaml and everything is set by default in there. During and after training, LoRAs will be saved in your outputs directory with the prefix _webui.

If you do not choose this setting, you will not currently be able to use these in the webui. If you train a Stable LoRA file, you cannot currently use them in inference.py.

Continue training a LoRA

To continue training a LoRA, simply set your lora_path in your config to the directory that contains your LoRA file(s), not an individual file. Each specific LoRA should have _unet or _text_encoder in the file name respectively, or else it will not work.

You should then be able to resume training from a LoRA model, regardless of which method you use (as long as the trained LoRA matches the version in the config).

What you cannot do:

• Use LoRA files that were made for SD image models in other trainers.
• Use 'cloneofsimo' LoRAs in another project (unless you build it or create a PR)
• Merge LoRA weights together (yet).

Finetune.

python train.py --config ./configs/v2/train_config.yaml

Training Results

With a lot of data, you can expect training results to show at roughly 2500 steps at a constant learning rate of 5e-6.

When finetuning on a single video, you should see results in half as many steps.

After training, you should see your results in your output directory.

By default, it should be placed at the script root under ./outputs/train_<date>

From my testing, I recommend:

• Keep the number of sample frames between 4-16. Use long frame generation for inference, not training.
• If you have a low VRAM system, you can try single frame training or just use n_sample_frames: 2.
• Using a learning rate of about 5e-6 seems to work well in all cases.
• The best quality will always come from training the text encoder. If you're limited on VRAM, disabling it can help.
• Leave some memory to avoid OOM when saving models during training.

Running inference

The inference.py script can be used to render videos with trained checkpoints.

Example usage:

python inference.py \
  --model camenduru/potat1 \
  --prompt "a fast moving fancy sports car" \
  --num-frames 60 \
  --window-size 12 \
  --width 1024 \
  --height 576 \
  --sdp

> python inference.py --help

usage: inference.py [-h] -m MODEL -p PROMPT [-n NEGATIVE_PROMPT] [-o OUTPUT_DIR]
                    [-B BATCH_SIZE] [-W WIDTH] [-H HEIGHT] [-T NUM_FRAMES]
                    [-WS WINDOW_SIZE] [-VB VAE_BATCH_SIZE] [-s NUM_STEPS]
                    [-g GUIDANCE_SCALE] [-i INIT_VIDEO] [-iw INIT_WEIGHT] [-f FPS]
                    [-d DEVICE] [-x] [-S] [-lP LORA_PATH] [-lR LORA_RANK] [-rw]

options:
  -h, --help            show this help message and exit
  -m MODEL, --model MODEL
                        HuggingFace repository or path to model checkpoint directory
  -p PROMPT, --prompt PROMPT
                        Text prompt to condition on
  -n NEGATIVE_PROMPT, --negative-prompt NEGATIVE_PROMPT
                        Text prompt to condition against
  -o OUTPUT_DIR, --output-dir OUTPUT_DIR
                        Directory to save output video to
  -B BATCH_SIZE, --batch-size BATCH_SIZE
                        Batch size for inference
  -W WIDTH, --width WIDTH
                        Width of output video
  -H HEIGHT, --height HEIGHT
                        Height of output video
  -T NUM_FRAMES, --num-frames NUM_FRAMES
                        Total number of frames to generate
  -WS WINDOW_SIZE, --window-size WINDOW_SIZE
                        Number of frames to process at once (defaults to full
                        sequence). When less than num_frames, a round robin diffusion
                        process is used to denoise the full sequence iteratively one
                        window at a time. Must be divide num_frames exactly!
  -VB VAE_BATCH_SIZE, --vae-batch-size VAE_BATCH_SIZE
                        Batch size for VAE encoding/decoding to/from latents (higher
                        values = faster inference, but more memory usage).
  -s NUM_STEPS, --num-steps NUM_STEPS
                        Number of diffusion steps to run per frame.
  -g GUIDANCE_SCALE, --guidance-scale GUIDANCE_SCALE
                        Scale for guidance loss (higher values = more guidance, but
                        possibly more artifacts).
  -i INIT_VIDEO, --init-video INIT_VIDEO
                        Path to video to initialize diffusion from (will be resized to
                        the specified num_frames, height, and width).
  -iw INIT_WEIGHT, --init-weight INIT_WEIGHT
                        Strength of visual effect of init_video on the output (lower
                        values adhere more closely to the text prompt, but have a less
                        recognizable init_video).
  -f FPS, --fps FPS     FPS of output video
  -d DEVICE, --device DEVICE
                        Device to run inference on (defaults to cuda).
  -x, --xformers        Use XFormers attnetion, a memory-efficient attention
                        implementation (requires `pip install xformers`).
  -S, --sdp             Use SDP attention, PyTorch's built-in memory-efficient
                        attention implementation.
  -lP LORA_PATH, --lora_path LORA_PATH
                        Path to Low Rank Adaptation checkpoint file (defaults to empty
                        string, which uses no LoRA).
  -lR LORA_RANK, --lora_rank LORA_RANK
                        Size of the LoRA checkpoint's projection matrix (defaults to
                        64).
  -rw, --remove-watermark
                        Post-process the videos with LAMA to inpaint ModelScope's
                        common watermarks.

Developing

Please feel free to open a pull request if you have a feature implementation or suggesstion! I welcome all contributions.

I've tried to make the code fairly modular so you can hack away, see how the code works, and what the implementations do.

Deprecation

If you want to use the V1 repository, you can use the branch here.

Shoutouts

• Showlab and bryandlee[https://github.com/bryandlee/Tune-A-Video] for their Tune-A-Video contribution that made this much easier.
• lucidrains for their implementations around video diffusion.
• cloneofsimo for their diffusers implementation of LoRA.
• kabachuha for their conversion scripts, training ideas, and webui works.
• JCBrouwer Inference implementations.
• sergiobr Helpful ideas and bug fixes.

Citation

If you find this work interesting, consider citing the original ModelScope Text-to-Video Technical Report:

@article{ModelScopeT2V,
  title={ModelScope Text-to-Video Technical Report},
  author={Wang, Jiuniu and Yuan, Hangjie and Chen, Dayou and Zhang, Yingya and Wang, Xiang and Zhang, Shiwei},
  journal={arXiv preprint arXiv:2308.06571},
  year={2023}
}