AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Layer-wise Vision Injection with Disentangled Attention for Efficient LVLMs

Xuange Zhang, Dengjie Li, Bo Liu, Zenghao Bao, Yao Zhou, Baisong Yang, Zhongying Liu, Yujie Zhong, Tongtong Yuan

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add layer-wise disentangled vision injection (LVIDA) to the decoderVicuna-7B LVLM · Decoder FLOPsGFLOPs: 10800 -> 1310Improved
add LVIDA to the decoderVicuna-7B LVLM · Time to first tokenTTFT: 215.6 -> 54.4Improved
add LVIDA to the decoderVicuna-7B LVLM · Peak memoryGB: 24.3 -> 17.2Improved
add LVIDA to the decoderVicuna-7B LVLM · MM-Vetaccuracy: 40.0 -> 37.5Hurt
add LVIDA to the decoderVicuna-7B LVLM · MMEscore: 1439.1 -> 1498.9Improved
add LVIDA to the decoderTinyLlama-1.1B LVLM · TextVQAaccuracy: 49.6 -> 44.1Hurt
add LVIDA instead of FastV or VTW visual token pruningLlama3-1B LVLM · TextVQAaccuracy: 38.7 -> 50.6Improved
add LVIDA instead of FastV (K=3,R=75%) pruningVicuna-7B LVLM · GQAaccuracy: 58.8 -> 62.4Improved
use a smaller TinyCLIP-39M encoder instead of SigLIP (ablation)LVIDA TinyLlama-1.1B · VQAv2accuracy: 75.5 -> 67.1Hurt
increase TinyCLIP feature dimension from 512 to 640 (40M to 61M) (ablation)LVIDA TinyLlama-1.1B · VQAv2accuracy: 64.2 -> 63.7No clear effect
use disentangled projection instead of vanilla projection (ablation)LVIDA Llama3-1B · GFLOPsGFLOPs: 2040 -> 192Improved
use disentangled projection instead of vanilla projection (ablation)LVIDA Llama3-1B · VQAv2accuracy: 76.8 -> 76.2No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy