MDM-2-DiffGAN implementation. Our paper

Set up dataset

HumanML3D - Follow the instructions in HumanML3D, then copy the result dataset to our repository:

cp -r ../HumanML3D/HumanML3D ./dataset/HumanML3D

Pretrained Weights

Download and save humanml-encoder-512 to save folder.

Training

To train our model, use the following script.

python -m train_ddgan --dataset humanml --num_channels 263 --batch_size 32

Generation

To generate a single prompt:

python -m sample.generate --dataset humanml --output_dir ./save/epoch325/toolbox --exp experiment --epoch_id 325 --text_prompt "the person walked forward and is picking up his toolbox."

To generate from test set prompts

python -m sample.generate --model_path ./save/humanml_trans_enc_512/model000200000.pt --num_samples 10 --num_repetitions 3

Evaluation

python -m eval.eval_humanml --dataset humanml --model_path saved_info/dd-gan/humanml/experiment/netG_325 --eval_mode mm_short --output_dir ./save --exp experiment --epoch_id 325 --node_rank 0 --text_prompt "A person jumping"

Acknowledgements

We want to thank "Human Motion Diffusion Model" and "TACKLING THE GENERATIVE LEARNING TRILEMMA WITH DENOISING DIFFUSION GANS" for their contributions. Their ideas, valuable insights, and codebase allowed us to implement our work.

License

Our code is distributed under both an MIT and NVIDIA License.