Research map / Vision-language and multimodal / CLIP and zero-shot
Inverse Optimal Transport for Efficient Adaptation of Vision-Language Models
Shupeng Qiu, Chuan-Xian Ren
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 |
|---|---|---|---|
| use IOTA instead of CLIP zero-shot | CLIP-RN50 / ViT-B/16 · 11 fine-grained datasets | accuracy (%): 56.04 -> 64.64 (RN50); 63.87 -> 71.73 (ViT-B/16) | Improved |
| use IOTA instead of graph-based TransCLIP/ECALP | CLIP-RN50 · 11 fine-grained datasets | accuracy (%): ECALP 64.00 -> IOTA 64.64 | Improved |
| use IOTA with uniform prior on imbalanced Caltech101 | CLIP-RN50 · Caltech101 | accuracy (%): 85.88 -> 76.84 | Hurt |
| use IOTA instead of ECALP on domain generalization | CLIP-RN50 · ImageNet variants | average accuracy (%): 47.73 -> 48.43 | Improved |
| add prototype adaptation to UOT (ablation) | IOTA-ZS · 11 fine-grained datasets | accuracy (%): 61.64 -> 63.32 (+1.68) | Improved |
| add adaptive threshold filtering (ablation) | IOTA-ZS · 11 fine-grained datasets | accuracy (%): 63.32 -> 64.32 (+1.0) | Improved |
| add label-guided mask in few-shot setting (ablation) | IOTA-FS · 16-shot, 11 datasets | accuracy (%): 65.25 -> 70.88 (+5.63) | Improved |
| use image-side-only marginal constraint instead of double constraint (ablation) | IOTA prototype learning · ImageNet | accuracy / runtime: accuracy 64.26 -> 64.25; runtime 56.76s -> 8.93s | No clear effect |
| vary UOT regularization weight lambda from 0.05 to 5 (ablation) | IOTA · fine-grained datasets | average accuracy (%): 64.25 to 64.37 | No clear effect |
| plug IOTA into CoOp prompt tuning | CoOp · 11 datasets, 1-16 shot | accuracy (%): up to +4.77% | Improved |
| plug IOTA into CLAP adapter | CLAP · 11 datasets, 1-16 shot | accuracy (%): up to +2.63% | Improved |
Closest papers by meaning
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.91)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- Efficient and Context-Aware Label Propagation for Zero-/Few-Shot Training-Free Adaptation of Vision-Language Model (ICLR 2025 · similarity 0.89)
- AWT: Transferring Vision-Language Models via Augmentation, Weighting, and Transportation (NeurIPS 2024 · similarity 0.89)
- Test-Time Spectrum-Aware Latent Steering for Zero-Shot Generalization in Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition (CVPR 2026 · 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.
