Research map / Vision-language and multimodal / CLIP and zero-shot
Hystar: Hypernetwork-driven Style-adaptive Retrieval via Dynamic SVD Modulation
Yujia Cai, Boxuan Li, Chenghao Xu, Jiexi Yan
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 hypernetwork + static modulation together over either alone (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 77.0 (static) / 70.6 (hyper) -> 85.3 | Improved |
| add StyleNCE loss instead of InfoNCE (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 85.3 -> 90.2 | Improved |
| add StyleNCE loss instead of InfoNCE (ablation) | Hystar on CLIP · DSR Low-Res query | Top-1 accuracy: 94.2 -> 98.0 | Improved |
| dynamic modulation on attention with static MLP rather than the reverse (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 71.2 -> 90.2 | Improved |
| dynamic modulation on attention only rather than on attention and MLP (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 69.9 -> 90.2 | Improved |
| inject hypernetwork at middle layers {4,7,10,13} instead of all layers 4-12 (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 76.5 -> 90.2 | Improved |
| inject at middle layers instead of only layer 1 (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 72.4 -> 90.2 | Improved |
| increase hypernetwork hidden width from 2048 to 4096 (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 90.2 -> 90.8 | No clear effect |
| add a third hypernetwork layer (depth 3 instead of 2) (ablation) | Hystar on CLIP · DSR Sketch query | Top-1 accuracy: 90.2 -> 90.4 | No clear effect |
| use DINOv2 vs CLIP(self) as style extractor (ablation) | Hystar on CLIP · DSR Top-1 average | Top-1 accuracy: 83.0 -> 83.6 | No clear effect |
| add style conditioning via hypernetwork over static-only modulation (ablation) | Hystar on CLIP · DSR Top-1 average | Top-1 accuracy: 73.9 -> 83.6 | Improved |
| Hystar vs FreestyleRet on inference cost | CLIP · Inference latency | speed (ms): 96 -> 108 ms; extra params 11.5% -> 3.5% | Hurt |
Closest papers by meaning
- FreestyleRet: Retrieving Images from Style-Diversified Queries (ECCV 2024 · similarity 0.87)
- StyleMeUp: Towards Style-Agnostic Sketch-Based Image Retrieval (CVPR 2021 · similarity 0.86)
- FashionERN: Enhance-and-Refine Network for Composed Fashion Image Retrieval (AAAI 2024 · similarity 0.85)
- ALADIN: All Layer Adaptive Instance Normalization for Fine-Grained Style Similarity (ICCV 2021 · similarity 0.85)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.84)
- Easy Regional Contrastive Learning of Expressive Fashion Representations (NeurIPS 2024 · similarity 0.84)
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.
