Research map / Vision-language and multimodal / CLIP and zero-shot
Toward a Holistic Evaluation of Robustness in CLIP Models
Weijie Tu, Weijian Deng, Tom Gedeon
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 |
|---|---|---|---|
| fine-tune on ImageNet-1K with cross-entropy | zero-shot CLIP ViT-B/32 · cue-conflict | shape bias: 0.575 -> 0.401 | Hurt |
| fine-tune on ImageNet-1K with cross-entropy | zero-shot CLIP ViT-B/16 · cue-conflict | shape bias: 0.473 -> 0.345 | Hurt |
| fine-tune with contrastive loss instead of cross-entropy (ablation) | CLIP ViT-B/32 · cue-conflict | shape bias: 0.401 -> 0.561 | Improved |
| fine-tune with contrastive loss instead of cross-entropy (ablation) | CLIP ViT-B/16 · cue-conflict | shape bias: 0.345 -> 0.448 | Improved |
| increase input resolution during fine-tuning (ablation) | CLIP · ImageNet val vs Stylized-ImageNet | accuracy | Mixed |
| additional fine-tuning on ImageNet-12K before ImageNet-1K (ablation) | CLIP-FT · OOD detection | AUROC | Improved |
| fine-tune CLIP on ImageNet | CLIP · calibration | ECE | Hurt |
| apply temperature scaling | zero-shot CLIP · ImageNet and shifts | NLL/ECE | Improved |
| use a single prompt instead of a large prompt set (ablation) | CLIP WIT models · classification, OOD, calibration | accuracy | Mixed |
| add LLM-generated prompt sets (VDT, MPVR, REAL) (ablation) | CLIP WIT models · classification | accuracy | Improved |
| swap CLIP for LLaVA with same visual encoder when category list is hard for CLIP | CLIP ViT-L/14-336 · ImageNet-D Background | accuracy: 0.14 -> 0.57 | Improved |
| swap CLIP for LLaVA with same visual encoder when category list is easy for CLIP | CLIP ViT-L/14-336 · ImageNet-D Background | accuracy: 0.90 -> 0.92 | No clear effect |
Closest papers by meaning
- What Makes CLIP More Robust to Long-Tailed Pre-Training Data? A Controlled Study for Transferable Insights (NeurIPS 2024 · similarity 0.91)
- Diffusion Feedback Helps CLIP See Better (ICLR 2025 · similarity 0.91)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.91)
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.91)
- A Sober Look at the Robustness of CLIPs to Spurious Features (NeurIPS 2024 · similarity 0.90)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.90)
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.
