Research map / Vision-language and multimodal / CLIP and zero-shot
Reevaluating the Intra-Modal Misalignment Hypothesis in CLIP
Jonas Herzog, Yue Wang
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 |
|---|---|---|---|
| replace OTI image-to-text inversion with PCA projection onto ImageNet class-name prompt subspace | image-image retrieval, CLIP ViT-B/32 · Retrieval, 13-dataset average | mAP: 42.9 -> 49.0 | Improved |
| replace OTI with PCA projection | image-image retrieval, SigLIP ViT-B/16 · Retrieval, 13-dataset average | mAP: 60.0 -> 62.8 | Improved |
| add PCA projection to raw image embeddings | CLIP ViT-L/14 retrieval · Retrieval, 13-dataset average | mAP: 53.7 -> 61.3 | Improved |
| add PCA projection to raw image embeddings | SigLIP and SigLIP2 retrieval · Retrieval, 13-dataset average | mAP: 57.2 -> 62.8 (SigLIP); 58.6 -> 64.4 (SigLIP2) | Improved |
| add PCA projection (information removal) for weather and time-of-day labels (ablation) | CLIP ViT-L/14 retrieval · BDD100k weather / time of day | mAP: -2.3 mAP (weather); -1.0 mAP (time of day) | Hurt |
| add image-image self-supervised objective (SigLIP2 vs SigLIP) (ablation) | SigLIP · Cosine similarity histograms, ImageNet val | class overlap and modality gap | No clear effect |
| add image-image objective (SigLIP2 vs SigLIP) | SigLIP retrieval · Retrieval, 13-dataset average | mAP: +1.2 mAP | Improved |
| add PCA projection to prototype classifier (ablation) | CLIP ViT-B/16 image-only few-shot · 11 datasets, 16-shot | accuracy: 73.5 -> 77.5 | Improved |
| add PCA projection to LDA classifier (ablation) | CLIP ViT-B/16 image-only few-shot · 11 datasets, 16-shot | accuracy: 79.5 -> 79.6 | No clear effect |
| add PCA projection to prototype classifier (ablation) | CLIP ViT-B/16 1-shot · 11 datasets, 1-shot | accuracy: 43.4 -> 50.7 | Improved |
| avoid image-image similarity (Tip-X, CODER) instead of image-space baselines | CLIP RN50 few-shot adaptation · 11 datasets few-shot | accuracy | No clear effect |
| swap CLIP for image-only DINO model | image-image classification Dogs vs Cats · Dogs vs Cats 1-shot | accuracy: 84.2 -> 76.2 (DINOv2) | Hurt |
Closest papers by meaning
- Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion (ICLR 2025 · similarity 0.95)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.90)
- Do Vision and Language Encoders Represent the World Similarly? (CVPR 2024 · similarity 0.90)
- SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 2025 · similarity 0.90)
- On the Value of Cross-Modal Misalignment in Multimodal Representation Learning (NeurIPS 2025 · similarity 0.89)
- Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models (ICLR 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.
