Research map / Vision-language and multimodal / CLIP and zero-shot
Soft Prompt Generation for Domain Generalization
Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, Badong Chen
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 fixed/conditional prompts with prompts generated by a CGAN (ablation) | CLIP RN50 prompt learning · Multi-source DG average of 5 datasets | accuracy: 66.3 -> 67.5 (vs CoCoOp) | Improved |
| use conditional prompts (CoCoOp, DPL) instead of a single fixed learned prompt (ablation) | CLIP RN50 prompt learning · Multi-source DG, TerraInc. | accuracy: 33.2 -> 34.1 | Improved |
| learn prompts instead of a manual prompt (ablation) | CLIP RN50 · TerraIncognita multi-source DG | accuracy: 23.8 -> 33.2 | Improved |
| average per-domain prompts (mix-domain) instead of one prompt from all domains (ablation) | fixed learned prompt · VLCS multi-source DG | accuracy: 81.4 -> 76.2 | Hurt |
| set prompt context length to 4 rather than 2, 8 or 16 (ablation) | SPG · PACS multi-source DG | average accuracy: 8: 88.3 -> 4: 92.8 | Improved |
| random Gaussian init instead of word-embedding init of the prompt (ablation) | SPG · PACS multi-source DG | average accuracy: 92.8 -> 92.2 | No clear effect |
| train the CGAN on more of the training data (ablation) | SPG · PACS multi-source DG | average accuracy: 30.7 (20%) -> 92.8 (full) | Improved |
| use SPG vs traditional DG methods on TerraIncognita | CLIP RN50 vs ResNet50 DG methods · TerraIncognita multi-source DG | accuracy: SPG 37.5 vs MIRO 50.4 | Hurt |
| use SPG in cross-dataset generalization (train on DomainNet) | CLIP ViT-B/16 · VLCS cross-dataset | accuracy: 80.5 (ZS-CLIP) -> 76.7 | Hurt |
| use SPG on cross-dataset generalization average | CLIP ViT-B/16 · Cross-dataset, 4 datasets average | accuracy: 72.0 -> 73.8 | Improved |
| use SPG in single-source DG | CLIP RN50 · Single-source DG, 5-dataset average | accuracy: 61.4 -> 62.8 (vs CoOp) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MaPLe — MaPLe: Multi-Modal Prompt Learning (CVPR 2023) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Domain Generalization in CLIP via Learning with Diverse Text Prompts (CVPR 2025 · similarity 0.89)
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.89)
- Domain-Agnostic Mutual Prompting for Unsupervised Domain Adaptation (CVPR 2024 · similarity 0.89)
- Concentrate Attention: Towards Domain-Generalizable Prompt Optimization for Language Models (NeurIPS 2024 · similarity 0.88)
- Concept-Guided Prompt Learning for Generalization in Vision-Language Models (AAAI 2024 · similarity 0.88)
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.
