CORTEXA
← Browse
arxivcs.CV2026-06-30

REDI: Corpus Aware Patch Ranking for DINOv3 Token Reduction

Chanjong Im, Sebastian Diem, Thomas Mandl

Most token reduction methods for Vision Transformers seek favorable tradeoffs between accuracy and efficiency by pruning, merging, or pooling patch tokens. REDI (Relevance for DINOv3 Token Reduction) studies this question through a controlled supervised reference: how should a fixed token budget be allocated across patches for image classification? REDI quantizes final block DINOv3 patch representations into a visual vocabulary and derives class conditioned corpus scores using supervised TF-IDF over visual words. For each validation image, the ground truth class selects a row of the TF-IDF table, and four transformed views produce a TF-IDF map aligned to a reference center crop. A separate dense pass on the same crop provides an attention map. After independent min max normalization, their elementwise product defines the REDI score. A fixed keep, merge, and compress operator then uses score rank to assign patch roles and score magnitude to weight merging and compression. With precomputed REDI scores, a frozen DINOv3 ViT-B/16 backbone, and the same linear classifier used for dense evaluation, the operator reduces the sequence length from 201 to 107 tokens, a 46.8% sequence reduction. The REDI variant based on incoming attention mass achieves 84.706% Top-1 accuracy on ImageNet-1K, compared with 83.514% for the dense baseline, 82.634% for incoming attention mass alone, and 81.796% for supervised TF-IDF alone. The same corpus term also improves reduced classification for three alternative attention formulations relative to their attention only counterparts. Together, these controlled comparisons indicate that class specific corpus statistics and image specific attention provide complementary signals for patch ranking in this setting.

View free PDFSource page

Related papers

arxivcs.CVcs.AI2026-07-06

TORINO: Token Reduction via Interpretable Concept Overlap in Vision-Language Models

Riccardo Renzulli, Gabriele Spadaro, Shruthi Gowda, Alaa Eddine Mazouz, Van-Tam Nguyen

Vision-Language Models (VLMs) have demonstrated impressive capabilities across different tasks, but their computational cost is dominated by the large number of visual tokens fed to the language model. Existing token reduction methods rely on attention-based scores or pairwise si…

View free PDFSource page
arxivcs.CV2026-07-17

Are All Tokens Necessary for Visual Place Recognition? An Empirical Study of Token Reduction for Efficient Inference

Tong Jin, Yunpeng Liu, Shuyu Hu, Qinghua Zhang, Ruize Han, Song Wang, et al.

Recent visual place recognition (VPR) methods based on vision transformers, particularly foundation models, have achieved remarkable recognition performance. However, these models process all visual tokens throughout the entire network, resulting in substantial computational over…

View free PDFSource page
arxivcs.CV2026-07-02

When Token Compression Breaks: Structural Pruning vs. Token Reduction for Robust ViT Segmentation under High Compression

Tien-Phat Nguyen, Ngai-Man Cheung

Vision Transformers (ViTs) are strong backbones for semantic segmentation, but their computational cost limits deployment. Recent token compression methods for efficient transformer-based segmentation reduce this cost by decreasing the number of tokens. However, existing evaluati…

View free PDFSource page
arxivcs.CV2026-07-15

Attention-Free and Lightweight Token Reduction for Efficient Vision-Language Models

Xuanyi Hao, Zuoyuan Zhang, Zhibo Wang, Xiaoyi Pang, Jiahui Hu, Jiacheng Du, et al.

Vision-Language Models (VLMs) have achieved strong performance in multimodal understanding, yet remain challenging to deploy on resource-constrained edge devices due to the substantial computational overhead of processing numerous visual tokens. Token reduction is a promising dir…

View free PDFSource page
arxivcs.CVcs.AI2026-07-24

TRaM-VSR: Importance-Aware Token Routing and Merging for One-Step Diffusion Video Super-Resolution

Sicheng Gao, Zhuyun Zhou, Yixuan Liu, Tong Shen, Zongwei Wu, Radu Timofte

Video super-resolution (VSR) using large-scale Diffusion Transformer (DiT) priors achieves exceptional perceptual quality but is often impractical due to the quadratic computational cost of processing dense spatio-temporal token sequences. Existing efficiency-oriented methods ris…

View free PDFSource page