This repository is the implementation of the Paper "3D-Aware Latent-Space Reenactment: Combining Expression Transfer and Semantic Editing".
Clone the repository.
git clone git@github.com:fraunhoferhhi/lasr3d.git
cd lasr_3d
Before obtaining the models, make sure that you have an Account to access the FLAME Model. We provided an install.sh script to build and configure the environment and obtain the models. It was tested with CUDA 11.7 and CUDA 11.8
source ./install.sh -a
conda activate lasr_3d
Alternatively, you can run ./install.sh -m to exclusively download the models, ./install.sh -e to exclusively set up the environment, or execute the individual steps in the installation script manually.
First, a model needs to be tuned on two individuals. Fine tuning the model takes about one hour of time.
python tune.py --id1 ./example/001.mp4 --id2 ./example/002.mp4 --out ./outputs/example/
Then, the tuned model can be used to edit the attributes of the people
python edit.py --model ./outputs/example/model.pkl --source 1 --target 0 --out ./outputs/example/ -e "glasses" +1.3 -e "sentiment" +0.3