facebookresearch/ToMe
Token Merging (ToMe) is a PyTorch implementation that accelerates Vision Transformers by merging similar tokens for 2-3x faster evaluation without retraining.
Not currently ranked — collecting fresh signals.
star history
ToMe provides a drop-in optimization for existing Vision Transformer architectures that merges redundant tokens based on similarity, effectively reducing computation while preserving accuracy. The method can be applied to pretrained ViT models without additional training and can further improve results when used during training. It also has extensions for diffusion models (ToMe-SD).
Frequently asked
- What is facebookresearch/ToMe?
- Token Merging (ToMe) is a PyTorch implementation that accelerates Vision Transformers by merging similar tokens for 2-3x faster evaluation without retraining.
- Is ToMe open source?
- Yes — facebookresearch/ToMe is an open-source project tracked on heatdrop.
- What language is ToMe written in?
- facebookresearch/ToMe is primarily written in Python.
- How popular is ToMe?
- facebookresearch/ToMe has 1.2k stars on GitHub.
- Where can I find ToMe?
- facebookresearch/ToMe is on GitHub at https://github.com/facebookresearch/ToMe.