Selective LoRA for Visual Tokens and Attention Heads

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Luo, Tiange, Logeswaran, Lajanugen, Kim, Jaekyeom, Johnson, Justin, Lee, Honglak
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910206743543808
author Luo, Tiange
Logeswaran, Lajanugen
Kim, Jaekyeom
Johnson, Justin
Lee, Honglak
author_facet Luo, Tiange
Logeswaran, Lajanugen
Kim, Jaekyeom
Johnson, Justin
Lee, Honglak
contents Low-rank adaptation (LoRA) is widely used for parameter-efficient fine-tuning, but its standard all-token, all-head design ignores the heterogeneous structure of vision language model (VLM) inputs. We introduce \emph{Image-LoRA}, a vision-oriented PEFT recipe that views LoRA as a token-level residual update and applies this update only to visual tokens. Image-LoRA further restricts adaptation to the value path of a compact subset of attention heads, selected using a one-pass influence estimate from a rank-1 visual-token-only probe. This token-, head-, and value-selective design reduces trainable parameters and adapter-only training FLOPs while leaving the pure-text forward pass of the frozen backbone unchanged when no visual tokens are present. Across visual localization benchmarks with controlled text:image token ratios, Image-LoRA matches or closely approaches standard LoRA, while showing especially favorable trade-offs in image-token-heavy regimes. We further validate its generality on TextVQA and VideoQA, verify pure-text preservation on GSM8K, and show on ViLP that a stronger information bottleneck can yield gains over standard LoRA.
format Preprint
id arxiv_https___arxiv_org_abs_2512_19219
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Selective LoRA for Visual Tokens and Attention Heads
Luo, Tiange
Logeswaran, Lajanugen
Kim, Jaekyeom
Johnson, Justin
Lee, Honglak
Computer Vision and Pattern Recognition
Artificial Intelligence
Low-rank adaptation (LoRA) is widely used for parameter-efficient fine-tuning, but its standard all-token, all-head design ignores the heterogeneous structure of vision language model (VLM) inputs. We introduce \emph{Image-LoRA}, a vision-oriented PEFT recipe that views LoRA as a token-level residual update and applies this update only to visual tokens. Image-LoRA further restricts adaptation to the value path of a compact subset of attention heads, selected using a one-pass influence estimate from a rank-1 visual-token-only probe. This token-, head-, and value-selective design reduces trainable parameters and adapter-only training FLOPs while leaving the pure-text forward pass of the frozen backbone unchanged when no visual tokens are present. Across visual localization benchmarks with controlled text:image token ratios, Image-LoRA matches or closely approaches standard LoRA, while showing especially favorable trade-offs in image-token-heavy regimes. We further validate its generality on TextVQA and VideoQA, verify pure-text preservation on GSM8K, and show on ViLP that a stronger information bottleneck can yield gains over standard LoRA.
title Selective LoRA for Visual Tokens and Attention Heads
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2512.19219