Introduction

This is the guidance for training your own MOFA-Adapter.

🔧🔧🔧 Stay tuned. There may exist bugs, feel free to contact me or report issues!

Environment Setup

conda create -n mofa_train python==3.10
conda activate mofa_train
pip install -r requirements.txt

Datset Preparation

We train our MOFA-Adapter on WebVid-10M. Please refer to our implementation of WebVid10M class in ./train_utils/dataset.py for more details about how to read the data. You may need to download the WebVid-10M first, or you can modify the codes of WebVid10M class and train your own MOFA-Adapter on other datasets.

Download checkpoints

1. Download the pretrained checkpoint folder of SVD_xt from huggingface to ./ckpts.

   The structure of the checkpoint folder should be:

   ./ckpts
   |-- stable-video-diffusion-img2vid-xt-1-1
   |   |-- feature_extractor
   |       |-- ...
   |   |-- image_encoder
   |       |-- ...
   |   |-- scheduler
   |       |-- ...
   |   |-- unet
   |       |-- ...
   |   |-- vae
   |       |-- ...
   |   |-- svd_xt_1_1.safetensors
   |   `-- model_index.json
   

2. Download the Unimatch checkpoint from here and put it into ./train_utils/unimatch/pretrained.

3. Download the checkpoint of CMP from here and put it into ./models/cmp/experiments/semiauto_annot/resnet50_vip+mpii_liteflow/checkpoints.

Run Training scripts

Stage 1

./train_stage1.sh

Stage 2

Change the value of --controlnet_model_name_or_path in train_stage2.sh, then run:

./train_stage2.sh