[ECCV2024] TCAN: Animating Human Images with Temporally Consistent Pose Guidance using Diffusion Models

This repository is the official implementation of TCAN

TCAN: Animating Human Images with Temporally Consistent Pose Guidance using Diffusion Models
Jeongho Kim*, Min-Jung Kim*, Junsoo Lee, Jaegul Choo (*: equal contribution)

[arXiv Paper]  [Project Page] 

🔖 TODO List

• Inference code
• Release model weights
• Training code

🗄️ Dataset

Preprocessed TikTok: Download

Unzip the donwnloaded dataset and set the path to the dataset as follows

cd TCAN
mkdir DATA
cd DATA
ln -s [data_path] TikTok 

TCAN/DATA/TikTok
L train
L valid_video

🌍 Environment

conda create -n tcan python=3.10
conda activate tcan
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2
pip install diffusers==0.25.0
pip install xformers==0.0.22
pip install accelerate==0.22.0
pip install transformers==4.32.0
pip install omegaconf
pip install einops
pip install clean-fid
pip install tensorboard
pip install imageio==2.9.0
pip install opencv-python
pip install av==11.0.0
pip install matplotlib
pip install peft==0.9.0
pip install imageio_ffmpeg
pip install ffmpeg
pip install scikit-image==0.20.0
pip install lpips
pip install onnxruntime
pip install numpy==1.26.4

⚖️ Model Weights

Pretrained Model Weights

📋 List of Pretrained Weights

• stablediffusion-v1.5
• VAE
• ControlNet

💻 Download From Terminal

git lfs install
cd checkpoints 
git clone https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5
git clone https://huggingface.co/stabilityai/sd-vae-ft-mse 
cd ..

# download yolox_l.onnx and dw-ll_ucoco_384.onnx
cd dwpose/annotator
git clone https://huggingface.co/yzd-v/DWPose ckpts

🔗 Download From Links

Place the downloaded weights into the 'TCAN/checkpoints' directory.

motion module weights provided by AnimateDiff: mm_sd_v15.ckpt, mm_sd_v15_v2.ckpt.

RealisticVision UNet weights: realisticVision

🔥 Train

🔥We trained our model using two A100🔥

1️⃣ First Stage

CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node 1 --master_port 3874 train.py \
 --config "./configs/train/first_stage.yaml" \
 --batch_size 2\
 --motion_type dwpose \
 --pretrained_unet_path "./checkpoints/realisticVisionV51_v20Novae.safetensors" \
 --pretrained_appearance_encoder_path "./checkpoints/realisticVisionV51_v20Novae.safetensors" \
 --pretrained_controlnet_path "./checkpoints/control_v11p_sd15_openpose_RenamedForMA.pth" \
 --freeze_controlnet \
 --init_unet_lora \
 --save_name First_Unetlora

2️⃣ Second Stage

CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node 1 --master_port 6836 train.py \
 --config ./configs/train/second_stage.yaml \
 --num_workers 2 \
 --batch_size 1 \
 --is_second_stage \
 --motion_type dwpose \
 --pretrained_unet_path "./logs/20240419_First_Unetlora/models/[UNet]_[Epoch=1]_[Iter=100]_[loss=0.1025].ckpt" \
 --pretrained_appearance_encoder_path "./logs/20240419_First_Unetlora/models/[AppearanceEncoder]_[Epoch=1]_[Iter=100]_[loss=0.1025].ckpt" \
 --pretrained_controlnet_path "./checkpoints/control_v11p_sd15_openpose_RenamedForMA.pth" \
 --init_unet_lora \
 --load_unet_lora_weight \
 --use_temporal_controlnet \
 --save_name SecondUnetloraTctrl