Research map / Vision-language and multimodal / CLIP and zero-shot
Decouple Before Align: Visual Disentanglement Enhances Prompt Tuning
Fei Zhang, Tianfei Zhou, Jiangchao Yao, Ya Zhang, Ivor Wai-Hung Tsang, Yanfeng Wang
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 visual triplet loss Lv (ablation) | DAPT-S baseline · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: +1.44 (novel -1.21) | Improved |
| add foreground-text alignment Lf (ablation) | DAPT-S baseline · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: +1.23 (novel -1.28) | Improved |
| add background-text alignment Lb (ablation) | DAPT-S baseline · StanfordCars+ImageNet avg, base-to-novel | novel accuracy: +1.60 novel, +1.05 16-shot | Improved |
| combine all three losses Lv, Lf, Lb (ablation) | DAPT-S baseline · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: 72.71 -> 75.22 | Improved |
| combine Lv and Lf without Lb (ablation) | DAPT-S baseline · StanfordCars+ImageNet avg, base-to-novel | novel accuracy: -2.24 | Hurt |
| use foreground-positive term only instead of full triplet (ablation) | Lv · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: +1.03 vs +1.44 | Hurt |
| use background-negative term only instead of full triplet (ablation) | Lv · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: +0.77 vs +1.44 | Hurt |
| replace hard filling with Gaussian blurring (ablation) | DAPT-S · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: 75.22 -> 74.09 | Hurt |
| increase background classes from 5 to 25 (ablation) | Lb only · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: 72.96 -> 73.76 (novel 72.17 -> 73.87) | Improved |
| use SEEM mask instead of Grad-CAM mask (ablation) | DAPT · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: 74.17 -> 75.22 | Improved |
| raise mask erasing rate from 0.1 to 0.7 (ablation) | DAPT-S · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: 75.22 -> 73.97 | Hurt |
| add DAPT to CoOp (ablation) | CoOp · StanfordCars+ImageNet avg, 16-shot | 16-shot accuracy: +2.85 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MaPLe — MaPLe: Multi-Modal Prompt Learning (CVPR 2023) · better on 5 of 5 shared results
- PromptKD — PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024) · better on 4 of 4 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
- Amend to Alignment: Decoupled Prompt Tuning for Mitigating Spurious Correlation in Vision-Language Models (ICML 2024 · similarity 0.91)
- Facing the Elephant in the Room: Visual Prompt Tuning or Full finetuning? (ICLR 2024 · similarity 0.89)
- Federated Disentangled Tuning with Textual Prior Decoupling and Visual Dynamic Adaptation (ICML 2025 · similarity 0.89)
- Revisiting the Power of Prompt for Visual Tuning (ICML 2024 · similarity 0.89)
- Learning to Learn Better Visual Prompts (AAAI 2024 · similarity 0.88)
- VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion (NeurIPS 2025 · 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.
