Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

OSLoPrompt: Bridging Low-Supervision Challenges and Open-Set Domain Generalization in CLIP

CVPR 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Fine-tuning strategies for few-shot · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces low-shot open-set domain generalization (LSOSDG) and OSLOPROMPT, a CLIP prompt-learning method with attribute cross-attention, visual prompts, and GPT-4o/Stable Diffusion fine-grained pseudo-open samples for an Unknown class.

Problem

Open-set domain generalization degrades in low-data regimes (1-shot and 5-shot), and existing methods struggle to reject open-set samples with fine-grained semantics close to training classes.

Approach

Learns domain-agnostic CLIP prompts regularized by domain-specific prompts enriched with GPT-4o-generated visual attributes via image-to-attribute cross-attention, plus shared learnable visual prompts; detects outliers with an Unknown-class prompt trained on fine-grained pseudo-open images generated by Stable Diffusion from GPT-4o-proposed near-class names.

Key results

• OSLOPROMPT achieves average H-score 75.57% (1-shot) and 76.21% (5-shot) across five datasets, beating ODG-CLIP by 11.93% and 9.01%.
• On ImageNet 1-shot closed-set DG, OSLOPROMPT averages 49.70% accuracy, 0.18% above the second-best method.
• Replacing fine-grained pseudo-open generation with the generic ODG-CLIP approach drops H-score by 22.95% (Office-Home) and 18.5% (Mini-DomainNet) in the ablation.

Measured effects reported in the paper

Surprising findings

• Generic pseudo-open generation from ODG-CLIP is far worse than the proposed fine-grained synthesis (41.09 vs 64.04 H-score on Office-Home 1-shot).
• Open-set H-score on Mini-DomainNet and Office-Home dips slightly from 1-shot to 5-shot, which the authors attribute to mild overfitting.

Limitations

• Relies on closed-source GPT-4o and Stable Diffusion for attribute and pseudo-open generation.
• Gains on open-set detection shrink or reverse in some 5-shot settings due to mild overfitting.
• No existing LSOSDG baselines, so competing methods were re-designed by the authors.

Experimental setup

Datasets: Office-Home PACS VLCS Mini-DomainNet Multi-Dataset ImageNet

Benchmarks: PACS VLCS Office-Home Multi-Dataset Mini-DomainNet ImageNet-V2 ImageNet-Sketch ImageNet-A ImageNet-R

Metrics: top-1 accuracy H-score

Baselines: ODG-CLIP StyLIP PromptSRC CLIPN CLIP + OpenMax MORGAN 2LM ODG-Net MEDIC SCI-PD CoOp CoCoOp

Builds on: CLIP ODG-CLIP CoOp StyLIP

Closest papers by meaning

• Unknown Prompt the only Lacuna: Unveiling CLIP's Potential for Open Domain Generalization (CVPR 2024 · similarity 0.94)
• Domain Generalization in CLIP via Learning with Diverse Text Prompts (CVPR 2025 · similarity 0.92)
• Learning to Adapt Frozen CLIP for Few-Shot Test-Time Domain Adaptation (ICLR 2025 · similarity 0.89)
• CLIPCEIL: Domain Generalization through CLIP via Channel rEfinement and Image-text aLignment (NeurIPS 2024 · similarity 0.89)
• Learning to Prompt with Text Only Supervision for Vision-Language Models (AAAI 2025 · similarity 0.89)
• PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.89)