ID-Selection: Prune 97% of Visual Tokens in Vision-Language Models While Keeping 92% of Performance

Less Is More: New Method Prunes 97% of Visual Tokens in LLaVA While Preserving Performance

Researchers have developed ID-Selection, a visual token selection strategy for Large Vision-Language Models (LVLMs) that achieves remarkable efficiency gains: pruning 97.2% of visual tokens while retaining only 16 tokens and maintaining 91.8% of original performance — all without additional training.

The Problem

Vision-language models like GPT-4V and LLaVA process images as sequences of visual tokens. Processing all tokens is extremely expensive. Existing pruning approaches face a trade-off:

• Importance-based → Retains redundant similar tokens
• Diversity-based → May discard informative tokens

ID-Selection's Innovation

The method couples importance with diversity in a unified process:

1. Score each visual token for importance
2. Select high-scoring tokens one by one
3. Progressively suppress similar tokens already represented

This ensures both informativeness and diversity without either dominating.

Results

Why This Matters

1. Cost reduction — Vision-language model inference can be 30x+ cheaper
2. Edge deployment — Makes powerful LVLMs feasible on mobile and edge devices
3. Speed improvement — Near-real-time vision understanding becomes practical
4. No retraining — Works with existing models out of the box
5. Energy efficiency — Critical for datacenter-scale vision AI