MakeupGrpoRL

Public asset bundle for the makeupGRPO project.

Contents

• rl_embeds_metadata_3200_relative.json: training metadata with paths relative to the dataset root. Use this as --data_json_path.
• rl_embeds_metadata_3200.json: original metadata with absolute source-machine paths, kept for provenance.
• archives/rl_processed_embeds_3200.zip: preprocessed UNO/FLUX conditioning tensors for 3,200 RL samples.
• archives/rl_makeup_final_v3.tar: makeup reference images used by the reward models.
• archives/rl_ref_bare_face.tar: bare-face reference images used by the reward models.
• archives/anchor_bare_face.tar: anchor bare-face image referenced by the original pre-encoding metadata.
• sft_lora/dev-bsz32_0303/checkpoint-20000/dit_lora.safetensors: current selected SFT LoRA starting checkpoint.
• face-parsing.PyTorch/: face parsing code and BiSeNet checkpoint used by the makeup extraction reward. The required checkpoint is face-parsing.PyTorch/res/cp/79999_iter.pth.

Usage Notes

For training in makeupGRPO, first extract the archives into one dataset root:

unzip archives/rl_processed_embeds_3200.zip
tar -xf archives/rl_makeup_final_v3.tar
tar -xf archives/rl_ref_bare_face.tar
tar -xf archives/anchor_bare_face.tar

Then set the dataset root to that directory and use:

--data_json_path rl_embeds_metadata_3200_relative.json
--sft_lora_path sft_lora/dev-bsz32_0303/checkpoint-20000/dit_lora.safetensors

The relative metadata is required because UNOPreprocessedDataset joins every tensor/image path with dataset_root.

The face parsing checkpoint is expected by the current code at a cluster-specific absolute path. Either mount or symlink this repo's face-parsing.PyTorch directory to that path, or update fastvideo/models/reward_model/makeup_sim_utils/face_parser_wrapper.py to point at:

face-parsing.PyTorch/res/cp/79999_iter.pth

The extracted directory should contain:

rl_embeds_metadata_3200_relative.json
rl_processed_embeds_3200/
rl_makeup_final_v3/
rl_ref_bare_face/
anchor_bare_face/