Research map / Vision-language and multimodal / CLIP and zero-shot
PowerCLIP: Powerset Alignment for Contrastive Pre-Training
Masaki Kawamura, Nakamasa Inoue, Rintaro Yanagi, Hirokatsu Kataoka, Rio Yokota
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 the triplet margin loss (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 35.1 -> 42.2 | Improved |
| Use region sets instead of individual regions (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 41.1 -> 42.2 | Improved |
| Use parse trees instead of individual tokens (ablation) | PowerCLIP-S · zero-shot retrieval average | average R@1 (%): 45.4 -> 47.0 | Improved |
| Add R2T aggregation (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 40.8 -> 42.2 | Improved |
| Add T2R aggregation (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 41.8 -> 42.2 | Improved |
| Use SAM masks instead of random masks (M=10) (ablation) | PowerCLIP · zero-shot classification and retrieval average | average top-1 accuracy (%): 41.5 -> 42.2 | Improved |
| Increase number of masks from 10 to 15 (ablation) | PowerCLIP · zero-shot retrieval average | average R@1 (%): 47.0 -> 44.9 (SAM) | Hurt |
| Use Softplus instead of ReLU in NLA-T1 (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 40.2 -> 42.2 | Improved |
| Use Tanh instead of Sigmoid in NLA-T2 (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 40.5 -> 42.2 | Improved |
| Use larger temperature 0.1 instead of 0.001 in NLAs (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 42.2 -> 40.7 | Hurt |
| Vary interpolation parameter in NLA-T2 from 0.75 to 0.50 (ablation) | PowerCLIP-S · zero-shot classification average | average top-1 accuracy (%): 42.2 -> 42.3 | No clear effect |
| Use SAM masks instead of random masks | PowerCLIP · Winoground group score | accuracy (%): 6.5 -> 10.2 | 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.
- SPARC — SPARC: Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models (CVPR 2025) · better on 2 of 2 shared results
- FILIP — FILIP: Fine-grained Interactive Language-Image Pre-Training (ICLR 2022) · better on 1 of 1 shared result
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
- SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 2025 · similarity 0.92)
- SuperCLIP: CLIP with Simple Classification Supervision (NeurIPS 2025 · similarity 0.91)
- Enhancing Compositional Reasoning in CLIP via Reconstruction and Alignment of Text Descriptions (NeurIPS 2025 · similarity 0.89)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.89)
- Mining Fine-Grained Image-Text Alignment for Zero-Shot Captioning via Text-Only Training (AAAI 2024 · similarity 0.89)
- TripletCLIP: Improving Compositional Reasoning of CLIP via Synthetic Vision-Language Negatives (NeurIPS 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.
