- Improve the dataset packing code.
- Add configuration for female dataset.
- Add the missing "ATUsquat" sequence to the male dataset configuration dictionary.
- Add performance report of our model on the two packed datasets.
- Add download links to various pretrained models, add config files of the pretrained models to the codebase.
-
Add scripts for training, test (auto-encoding) and data preprocessing.
-
Add feature: a new type of residual block for the decoder (
res_block_affine), based on the idea of this CVPR2020 paper. It achieves improved performance than our original mesh res-block. To use this layer use the flag--affine 1; otherwise it uses our graph res-block by default. -
Add feature: two new model configurations with 4 layers or 6 layers respectively. Specify this with the flag e.g.
--num_conv_layers 6 -
Add feature: now the spatial downsampling factor can be set by
--ds_factor X(previously hard-set as 2).
- Fix the download link to a faulty checkpoint that yields strange generated garment shape.
- Fix a bug that causes the group norm stats not properly loaded at inference.