Research map / Vision-language and multimodal / CLIP and zero-shot
Improving CLIP Training with Language Rewrites
Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, Yonglong Tian
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 LLaMA ICL caption rewrites as text augmentation | CLIP ViT-B/16 · CC12M ImageNet zero-shot | top-1 accuracy: 40.2 -> 48.4 | Improved |
| add LLM rewrites on LAION-400M | CLIP ViT-B/32 · LAION-400M ImageNet zero-shot | top-1 accuracy: 62.0 -> 64.4 | Improved |
| add LLM rewrites to SLIP (LaSLIP) | SLIP · CC12M ImageNet zero-shot | top-1 accuracy: 42.1 -> 49.7 | Improved |
| add LLM rewrites to CLIP | CLIP · CC12M few-shot average | 5-way 5-shot accuracy: 83.3 -> 85.8 | Improved |
| add LLM rewrites to CLIP | CLIP · CC12M ImageNet linear probe | accuracy: 70.3 -> 72.3 | Improved |
| use EDA word-level augmentation instead of LLM rewrites (ablation) | LaCLIP · CC12M ImageNet zero-shot | top-1 accuracy: 48.4 -> 41.2 | Hurt |
| use back translation instead of LLM rewrites (ablation) | LaCLIP · CC12M ImageNet zero-shot | top-1 accuracy: 48.4 -> 41.6 | Hurt |
| vary source of meta-input-output ICL examples (ChatGPT/Bard/COCO/Human) (ablation) | LaCLIP · CC12M ImageNet zero-shot | top-1 accuracy: 44.5 / 44.8 / 44.6 / 45.1 | No clear effect |
| increase number of LLM rewrites per caption (ablation) | LaCLIP · CC12M ImageNet zero-shot | top-1 accuracy | Improved |
| use LaCLIP across ViT-S/16, B/16, L/16 (ablation) | CLIP · CC12M ImageNet zero-shot | top-1 accuracy: 36.9 -> 46.3 (S), 44.0 -> 49.1 (L) | Improved |
| initialize and finetune text encoder from BERT instead of LaCLIP rewrites (ablation) | CLIP · CC12M ImageNet zero-shot | top-1 accuracy: 40.2 -> 42.9 | Improved |
| freeze pretrained BERT text encoder (ablation) | CLIP · CC12M ImageNet zero-shot | top-1 accuracy: 40.2 -> 23.2 | Hurt |
Closest papers by meaning
- Diffusion Feedback Helps CLIP See Better (ICLR 2025 · similarity 0.92)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.92)
- Meta CLIP 2: A Worldwide Scaling Recipe (NeurIPS 2025 · similarity 0.92)
- AmorLIP: Efficient Language-Image Pretraining via Amortization (NeurIPS 2025 · similarity 0.92)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.92)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.91)
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.
