Meanwhile, ByteDance also released a paper named MagicAnimate. The paper can be viewed here: https://arxiv.org/abs/2311.16498.
ByteDance also open-sourced the code for MagicAnimate: https://github.com/magic-research/magic-animate.
Effect demonstration:
The results of MagicAnimate are quite impressive.
Below are some examples and comparisons with other algorithms.
Brief introduction to technical principles:
The working principle of MagicAnimate is shown in the figure below.
Based on a reference image and target DensePose motion sequence, MagicAnimate uses a video diffusion model and appearance encoder for temporal modeling and identity preservation (on the left). To support long video animations, we designed a simple video fusion strategy that produces smooth video transitions during inference (on the right).
Usage on Hugging Face:
You can try MagicAnimate on the Hugging Face platform via the following link: https://huggingface.co/spaces/zcxu-eric/magicanimate.
However, when I tried it, I couldn't get through due to too many users. I attempted running it myself on Google Colab:
%cd /content
!git clone -b dev https://github.com/camenduru/MagicAnimate-hf
!pip install -q diffusers==0.21.4 transformers==4.32.0 accelerate==0.22.0 omegaconf==2.3.0 einops==0.6.1 av gradio
!pip install -q https://download.pytorch.org/whl/cu118/xformers-0.0.22.post4%2Bcu118-cp310-cp310-manylinux2014_x86_64.whl
!apt -y install -qq aria2
BaseModelUrl = "https://huggingface.co/runwayml/stable-diffusion-v1-5"
BaseModelDir = "/content/MagicAnimate-hf/stable-diffusion-v1-5"
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/model_index.json -d {BaseModelDir} -o model_index.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/resolve/main/vae/diffusion_pytorch_model.bin -d {BaseModelDir}/vae -o diffusion_pytorch_model.bin
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/vae/config.json -d {BaseModelDir}/vae -o config.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/resolve/main/unet/diffusion_pytorch_model.bin -d {BaseModelDir}/unet -o diffusion_pytorch_model.bin
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/unet/config.json -d {BaseModelDir}/unet -o config.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/tokenizer/vocab.json -d {BaseModelDir}/tokenizer -o vocab.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/tokenizer/tokenizer_config.json -d {BaseModelDir}/tokenizer -o tokenizer_config.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/tokenizer/special_tokens_map.json -d {BaseModelDir}/tokenizer -o special_tokens_map.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/tokenizer/merges.txt -d {BaseModelDir}/tokenizer -o merges.txt
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/resolve/main/text_encoder/pytorch_model.bin -d {BaseModelDir}/text_encoder -o pytorch_model.bin
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/text_encoder/config.json -d {BaseModelDir}/text_encoder -o config.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/scheduler/scheduler_config.json -d {BaseModelDir}/scheduler -o scheduler_config.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/resolve/main/safety_checker/pytorch_model.bin -d {BaseModelDir}/safety_checker -o pytorch_model.bin
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/safety_checker/config.json -d {BaseModelDir}/safety_checker -o config.json
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M {BaseModelUrl}/raw/main/feature_extractor/preprocessor_config.json -d {BaseModelDir}/feature_extractor -o preprocessor_config.json
%cd /content/MagicAnimate-hf
!python app.py
Check out the resultsScreenshot
