Busan Token-pruning Series
CVSP mascot Busan · Gwangan Bridge at night Boogi — Busan mascot

BTS is a curated series of token & KV-cache pruning methods that ⚡ prune away redundancy while preserving performance across image, 3D, and text — on a mission to become the first pruning series with a method for every modality, from the PNU-CVSP lab.

stars modalities
Overview

One goal, every modality

The Busan Token-pruning Series is not confined to a single modality. It shares one goal — ⚡ prune redundant tokens for efficient inference — while each work develops its own method for a different modality: image, 3D, and text. A common goal, modality-specific solutions.

The Series

⚡ Projects

🖼️

AgilePruner

ImageICLR 2026

Efficient visual token pruning for 2D vision models — jointly analyzes token complexity, diversity, and attention to drop redundant tokens while keeping accuracy.

🧊

3DZip

3DECCV 2026

Spatial-aware, feature-diversity-guided token compression for 3D question answering — retains 94.7% performance with only 128 tokens at 1.92× faster inference.

💬

HybridKV

TextICMLw 2026Query-agnostic

Query-agnostic KV-cache compression via head-dominant context reconstruction — matches full reconstruction at up to 36% less compression overhead.

💬

RestoreKV NEW

TextPreprintQuery-agnostic

Complements query-agnostic KV cache eviction with a learned, budget-matched restoration pass — recovering full-cache behavior under aggressive compression while training only 0.4% of parameters.