Research map / Architectures and efficiency / Attention and transformers
LookupViT: Compressing visual information to a limited number of tokens
Rajat Koner, Gagan Jain, Sujoy Paul, Volker Tresp, Prateek Jain
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add bidirectional information flow (MHBC compressed-to-lookup) to LookupViT (ablation) | LookupViT B/16 5x5 · ImageNet-1k (from scratch) | top-1 accuracy: 69.7 -> 79.1 | Improved |
| add lookup tokens with cross-attention to compressed tokens (ablation) | LookupViT B/16 5x5 · ImageNet-1k (from scratch) | top-1 accuracy: 69.2 -> 79.1 | Improved |
| add classification loss on lookup tokens (ablation) | LookupViT B/16 5x5 · ImageNet-1k (from scratch) | top-1 accuracy: 78.2 -> 79.1 | Improved |
| add classification loss on compressed tokens (ablation) | LookupViT B/16 5x5 · ImageNet-1k (from scratch) | top-1 accuracy: 78.4 -> 79.1 | Improved |
| build compressed tokens by parameter-free resize of lookup tokens instead of random learnable tokens (ablation) | LookupViT B/16 5x5 · ImageNet-1k (from scratch) | top-1 accuracy: 77.9 -> 79.1 | Improved |
| replace ViT with LookupViT 5x5 (B/16, from scratch) | ViT B/16 · ImageNet-1k (from scratch) | top-1 accuracy and GFLOPs: 78.6 -> 79.1 accuracy, 35.1 -> 16.5 GFLOPs | Improved |
| replace ViT with LookupViT 3x3 after IN-21k pretraining | ViT B/16 · ImageNet-1k (IN-21k pretrain, finetune) | top-1 accuracy: 83.7 -> 77.9 | Hurt |
| use LookupViT instead of ViT on corrupted/OOD images | ViT B/16 · ImageNet-C | accuracy: 56.4 -> 59.9 | Improved |
| use LookupViT 5x5 instead of ViT on ImageNet-A | ViT B/16 · ImageNet-A | accuracy: 6.7 -> 5.4 | Hurt |
| use LookupViViT instead of ViViT on SSv2 at half FLOPs | ViViT-B · Something-Something V2 (1 view) | accuracy: 51.2 -> 56.3 | Improved |
| use LookupViViT instead of ViViT on Kinetics400 | ViViT-B · Kinetics400 (12 views) | accuracy: 78.6 -> 77.9 | No clear effect |
| train with random compressed-token resolution (Multi-Res) | LookupViT · ImageNet-1k (IN-21k pretrain, finetune) | top-1 accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Perceiver — Perceiver: General Perception with Iterative Attention (ICML 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Making Vision Transformers Efficient From a Token Sparsification View (CVPR 2023 · similarity 0.91)
- DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification (NeurIPS 2021 · similarity 0.89)
- A-ViT: Adaptive Tokens for Efficient Vision Transformer (CVPR 2022 · similarity 0.89)
- SparseViT: Revisiting Activation Sparsity for Efficient High-Resolution Vision Transformer (CVPR 2023 · similarity 0.89)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.89)
- LookWhere? Efficient Visual Recognition by Learning Where to Look and What to See from Self-Supervision (NeurIPS 2025 · similarity 0.89)
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.
