Research map / Vision-language and multimodal / CLIP and zero-shot
Self-regulating Prompts: Foundational Model Adaptation without Forgetting
Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, Fahad Shahbaz Khan
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 |
|---|---|---|---|
| add self-consistency loss (mutual agreement with frozen CLIP features) (ablation) | IVLP · Base-to-novel, 11 datasets avg | Novel accuracy: 71.79 -> 75.38 | Improved |
| add Gaussian weighted prompt aggregation (ablation) | IVLP + SCL · Base-to-novel, 11 datasets avg | HM: 79.55 -> 79.70 | Improved |
| add textual diversity with multiple templates (ablation) | IVLP + SCL + GPA · Base-to-novel, 11 datasets avg | HM: 79.70 -> 79.97 | Improved |
| use L1 instead of cosine similarity or MSE for feature matching (ablation) | SCL · Base-to-novel, 11 datasets avg | HM: 79.17 (cosine), 79.33 (MSE) -> 79.43 (L1) | Improved |
| use Gaussian weighting instead of equal weights or EMA for prompt aggregation (ablation) | PromptSRC · Base-to-novel, 11 datasets avg | HM: 79.83 (equal), 79.47 (EMA) -> 79.97 | Improved |
| increase number of text templates for textual diversity (ablation) | PromptSRC · ImageNet | HM | Improved |
| increase prompt length (ablation) | PromptSRC · Base-to-novel | HM | Improved |
| PromptSRC instead of MaPLe on base-to-novel | CLIP ViT-B/16 prompting · Base-to-novel, 11 datasets avg | HM: 78.55 -> 79.97 | Improved |
| PromptSRC instead of MaPLe on cross-dataset transfer | CLIP ViT-B/16 prompting · Cross-dataset target avg | accuracy: 66.30 -> 65.81 | Hurt |
| PromptSRC instead of MaPLe on domain generalization | CLIP ViT-B/16 prompting · Domain generalization target avg | accuracy: 60.27 -> 60.65 | Improved |
| add regularization training overhead (extra image forward pass) | IVLP · SUN397 compute | Train time (min): 12.01 -> 13.13 | Hurt |
| apply PromptSRC to video model ViFi-CLIP instead of IVLP | ViFi-CLIP · UCF-101 video base-to-novel | HM: 83.60 -> 85.50 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MaPLe — MaPLe: Multi-Modal Prompt Learning (CVPR 2023) · better on 5 of 7 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
- VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion (NeurIPS 2025 · similarity 0.90)
- Revisiting the Power of Prompt for Visual Tuning (ICML 2024 · similarity 0.90)
- A Similarity Paradigm Through Textual Regularization Without Forgetting (AAAI 2025 · similarity 0.89)
- Aggregate-and-Adapt Natural Language Prompts for Downstream Generalization of CLIP (NeurIPS 2024 · similarity 0.89)
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Learning to Learn Better Visual Prompts (AAAI 2024 · 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.
