Research map / Vision-language and multimodal / CLIP and zero-shot
A Sentence Speaks a Thousand Images: Domain Generalization through Distilling CLIP with Language Guidance
Zeyi Huang, Andy Zhou, Zijian Ling, Mu Cai, Haohan Wang, Yong Jae Lee
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 CLIP distillation hint (KL on teacher logits) to ERM (ablation) | ResNet50 · PACS | accuracy: 85.7 -> 88.4 | Improved |
| add absolute distance loss to ERM + Hint (ablation) | ResNet50 · PACS | accuracy: 88.4 -> 89.0 | Improved |
| add relative distance loss on top of Hint + AD (ablation) | ResNet50 · PACS | accuracy: 89.0 -> 89.4 | Improved |
| add AD + RD to ERM + Hint | ResNet50 · DomainBed average | average accuracy: 72.5 -> 73.8 | Improved |
| add AD + RD to ERM + Hint | ResNet18 · DomainBed average | average accuracy: 68.6 -> 69.6 | Improved |
| use CLIP text embedding instead of image embedding for AD supervision (ablation) | ERM + Hint + AD · PACS | accuracy: 88.4 -> 89.0 | Improved |
| use 80-template prompt ensemble instead of a single prompt (ablation) | ERM + Hint + AD · PACS | accuracy: 88.5 -> 89.0 | Improved |
| use cosine similarity for absolute distance instead of L1/L2/supervised contrastive (ablation) | ERM + Hint + AD · PACS | accuracy: 88.0 / 88.1 / 88.6 -> 89.0 | Improved |
| use L2 vs KL for k1 in relative distance loss (ablation) | ERM + Hint + RD · PACS | accuracy: 88.7 -> 88.8 | No clear effect |
| AD alone (text embedding only, no distillation hint) vs distillation hint (ablation) | ERM · PACS | accuracy: 88.4 (Hint) vs 87.8 (AD) | No clear effect |
| ensemble of students distilled from two different CLIP teachers (Mix Teacher) (ablation) | RISE ResNet50 · PACS | accuracy: 89.4 -> 90.2 | Improved |
| ensemble two students from the same ViT B/16 teacher (ablation) | RISE ResNet50 · PACS | accuracy: 89.4 -> 89.8 | Improved |
Closest papers by meaning
- Domain Generalization in CLIP via Learning with Diverse Text Prompts (CVPR 2025 · similarity 0.93)
- CLIPCEIL: Domain Generalization through CLIP via Channel rEfinement and Image-text aLignment (NeurIPS 2024 · similarity 0.93)
- Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification (CVPR 2024 · similarity 0.92)
- When and How Does CLIP Enable Domain and Compositional Generalization? (ICML 2025 · similarity 0.90)
- Unknown Prompt the only Lacuna: Unveiling CLIP's Potential for Open Domain Generalization (CVPR 2024 · similarity 0.90)
- A Simple Recipe for Language-guided Domain Generalized Segmentation (CVPR 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.
