ERASE: Eliminating Redundant Visual Tokens via Adaptive Two-Stage Token Pruning

Name: ERASE: Eliminating Redundant Visual Tokens via Adaptive Two-Stage Token Pruning
Uploaded: 2026-10-11T06:04:51.171761+00:00
Duration: 274 s
Description: ERASE is an adaptive two-stage framework to compress vision tokens in Vision-Language Models for better efficiency. It uses image entropy to decide how much visual redundancy to remove in Stage 1, and then uses context-aware attention scores in Stage 2. This dynamic approach ensures high accuracy retention while significantly speeding up inference on high-resolution images.

ERASE is an adaptive two-stage framework to compress vision tokens in Vision-Language Models for better efficiency. It uses image entropy to decide how much visual redundancy to remove in Stage 1, and then uses context-aware attention scores in Stage 2. This dynamic approach ensures high accuracy retention while significantly speeding up inference on high-resolution images.

Open the full episode