Research map / Vision-language and multimodal / CLIP and zero-shot
Beyond Words: Augmenting Discriminative Richness via Diffusions in Unsupervised Prompt Learning
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.
Proposes AiR, which fine-tunes Stable Diffusion with LoRA on confidently pseudo-labeled images, generates synthetic class images, and uses them as an auxiliary image-image classifier fused with CLIP's text classifier to produce better pseudo-labels and extra training samples for unsupervised prompt learning.
Problem
Pseudo-labels produced by CLIP text classifiers in unsupervised prompt learning are often inaccurate because class-name text cannot capture visual variation, leading to sub-optimal prompt tuning.
Approach
Use LoRA-fine-tuned Stable Diffusion to generate synthetic class images, select the most CLIP-aligned synthetic image per class as visual prototypes for an auxiliary classifier, and fuse its predictions with the text classifier for pseudo-labeling, while also training prompts on synthetic images.
Key results
- AiR improves text prompt learning by about 2.6% average accuracy over the prior SOTA CPL across five datasets in the UL setting.
- AiR improves visual prompt learning by about 2.7% average over CPL in UL.
- Pseudo-label accuracy on RESISC45 improves by about 6% over CLIP text-only classification on the top-50 confident samples.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add AiR's diffusion-generated image classifier and synthetic-image training to CPL | text prompt learning with CLIP ViT-B/32 · RESISC45, FGVCAircraft, Flowers102, EuroSAT, and DTD; SSL, UL, and TRZSL | top-1 test accuracy (%): 82.5 vs 80.9; 79.9 vs 77.3; 87.5 vs 85.8 (RESISC45) | Helped |
| add AiR's diffusion-generated image classifier and synthetic-image training to CPL | text prompt learning with CLIP ViT-B/32 · FGVCAircraft; SSL, UL, and TRZSL | top-1 test accuracy (%): 24.2 vs 22.4; 19.6 vs 18.3; 33.2 vs 30.8 | Helped |
| add AiR's diffusion-generated image classifier and synthetic-image training to CPL | text prompt learning with CLIP ViT-B/32 · Flowers102; SSL, UL, and TRZSL | top-1 test accuracy (%): 91.2 vs 89.6; 74.3 vs 72.9; 92.3 vs 87.3 | Helped |
| add AiR's diffusion-generated image classifier and synthetic-image training to CPL | text prompt learning with CLIP ViT-B/32 · EuroSAT; SSL, UL, and TRZSL | top-1 test accuracy (%): 76.4 vs 75.4; 71.4 vs 67.2; 93.9 vs 93.7 | Helped |
| add AiR's diffusion-generated image classifier and synthetic-image training to CPL | text prompt learning with CLIP ViT-B/32 · DTD; SSL, UL, and TRZSL | top-1 test accuracy (%): 62.2 vs 61.2; 55.7 vs 51.9; 69.9 vs 68.0 | Helped |
| add LoRA fine-tuning of Stable Diffusion using pseudo-labeled data (ablation) | AiR · RESISC45 UL | top-1 test accuracy (%): 76.5 -> 79.9+3.4 | Helped |
| add LoRA fine-tuning of Stable Diffusion using pseudo-labeled data (ablation) | AiR · Flowers102 UL | top-1 test accuracy (%): 66.3 -> 74.3+8.0 | Helped |
| add LoRA fine-tuning of Stable Diffusion using pseudo-labeled data (ablation) | AiR · EuroSAT UL | top-1 test accuracy (%): 66.2 -> 71.4+5.2 | Helped |
| add LoRA fine-tuning of Stable Diffusion using pseudo-labeled data (ablation) | AiR · DTD UL | top-1 test accuracy (%): 51.6 -> 55.7+4.1 | Helped |
| add the synthetic-sample auxiliary classifier (use its predictions alongside real-sample loss) (ablation) | GRIP on RESISC45, Flowers102, and DTD · RESISC45, Flowers102, and DTD UL | accuracy (%): 70.6 -> 72.3; 69.8 -> 71.6; 46.0 -> 51.4 | Helped |
| add synthetic-sample training loss to real-sample training loss (ablation) | GRIP on RESISC45, Flowers102, and DTD · RESISC45, Flowers102, and DTD UL | accuracy (%): 70.6 -> 72.9; 69.8 -> 71.3; 46.0 -> 50.8 | Helped |
| train using synthetic samples only, without real-sample loss (ablation) | AiR · RESISC45 and Flowers102 UL | accuracy (%): RESISC45: 73.6 to 58.3; Flowers102: 71.9 to 59.7 | Hurt |
Surprising findings
- Training on synthetic samples alone degrades accuracy sharply (e.g., -10.1% on Flowers102 and -12.3% on RESISC45), so real samples remain necessary.
Limitations
- Relies on synthetic samples alone being insufficient, so real unlabeled data remains essential.
- Gains plateau beyond about 120 synthetic samples per class, indicating a limit to the benefit.
- Selection of the representative synthetic image uses CLIP zero-shot similarity, which depends on the quality of the text classifier.
Experimental setup
Datasets: DTD EuroSAT FGVC-Aircraft Flowers102 RESISC45
Benchmarks: DTD EuroSAT FGVC-Aircraft Flowers102 RESISC45
Metrics: top-1 accuracy harmonic mean accuracy (TRZSL)
Baselines: CLIP zero-shot CoOp VPT FPL GRIP CPL
Builds on: CLIP CPL GRIP DiffTPT CoOp
Closest papers by meaning
- ProAPO: Progressively Automatic Prompt Optimization for Visual Classification (CVPR 2025 · similarity 0.90)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.89)
- Prompting Hard or Hardly Prompting: Prompt Inversion for Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.89)
- NLPrompt: Noise-Label Prompt Learning for Vision-Language Models (CVPR 2025 · similarity 0.89)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.88)
- CoPL: Contextual Prompt Learning for Vision-Language Understanding (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.