FlowCut
Information-flow-based token pruning for efficient VLMs.
FlowCut rethinks visual token redundancy from the perspective of information flow across Transformer layers.
It dynamically selects tokens according to flow-based importance, enabling high pruning ratios and faster inference while maintaining performance on multimodal understanding tasks.
- Paper: arXiv:2505.19536
- Code: TungChintao/FlowCut