Request for Pre-trained Tokenizer Weights (pytorch_model.bin) for Reproducibility #3
Description
First of all, thank you for sharing your amazing work! The framework is impressive.
While exploring the codebase and the training configurations, I noticed that the config.py files for all stages refer to local paths for the image tokenizer checkpoints, specifically:
pytorch_model.bin (referenced in Stage 1 & 2)
pytorch_model_512.bin (referenced in Stage 3)（which can be found in huggingface）
The paths in the config (e.g., /opt/tiger/ju/ckpt/flowtitok_swiglu_bl77_vae/...) suggest these are internal checkpoints. Since the image tokenizer is the "foundation" for the Flow Matching training, it is currently difficult for the community to:
Reproduce the Stage 1/2 pre-training results.
Could you kindly consider sharing these tokenizer weights (e.g., on Hugging Face or GitHub Releases)?