SFPruner

[ECCV 2026] SFPruner: Structured Redundancy Modeling for Efficient Visual Token Pruning in High-Resolution MLLMs

Juwon Song1, Woohyeong Kim2, Kyeongbo Kong2
1LG Electronics    2Pusan National University

Project page     arXiv: coming soon     Code: coming soon

SFPruner Algorithm

SFPruner pipeline

SFPruner consists of three stages:

  1. Semantic guidance
    Estimate base token importance using instruction relevance and visual saliency.

  2. Semantics-Guided Ridge Leverage Score (SG-RLS)
    Suppress globally redundant covariance directions by reweighting token scores with ridge leverage-based structural information.

  3. Ranking-Based Directional Masking
    Resolve residual pairwise redundancy by allowing higher-scoring tokens to suppress similar lower-scoring alternatives through a single parallel masking operation.

This design avoids the iterative dependency of subset-optimization methods such as DPP-, diversity-, or graph-based selection, enabling direct Top-K pruning with stable selection latency.