Research map / Architectures and efficiency / Pruning and efficient tuning
LoCA: Spatially-Aware Low-Rank Convolutional Adaptation of Vision Foundation Models
Sojung An, Junha Lee, Sujeong You, Nam Ik Cho, Donghyun Kim
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 |
|---|---|---|---|
| apply flattened convolutional LoRA instead of linear-only LoRA (ablation) | LoRA on ConvNeXt-L · GTAV to Cityscapes/BDD/Mapillary DGSS | average mIoU: 65.29 -> 64.95 | Hurt |
| add low-rank channel mixing to LoRA baseline (ablation) | LoCA on ConvNeXt-L · GTAV to Cityscapes/BDD/Mapillary DGSS | average mIoU: 65.29 -> 66.16 | Improved |
| add SVD-based spatial basis refinement (ablation) | LoCA on ConvNeXt-L · GTAV to Cityscapes/BDD/Mapillary DGSS | average mIoU: 66.16 -> 67.03 | Improved |
| add hierarchical rank scheduling (ablation) | LoCA on ConvNeXt-L · GTAV to Cityscapes/BDD/Mapillary DGSS | average mIoU: 67.03 -> 67.46 | Improved |
| use covariance-SVD initialization instead of zero initialization (ablation) | LoCA · VTAB-1k average | average accuracy: 75.7 -> 75.9 | Improved |
| use covariance-SVD initialization instead of flattened SVD (ablation) | LoCA · VTAB-1k average | average accuracy: 73.0 -> 75.9 | Improved |
| use covariance-SVD initialization instead of zero initialization (ablation) | LoCA · DGSS average | average mIoU: 65.6 -> 66.5 | Improved |
| initialize channel mixing with SoMA principal singular components instead of zero (ablation) | LoCA on ConvNeXt-L · GTAV to Cityscapes/BDD/Mapillary DGSS | average mIoU: 66.49 -> 67.46 | Improved |
| initialize channel mixing with SoMA principal singular components instead of zero (ablation) | LoCA on ConvNeXt-B · GTAV to Cityscapes/BDD/Mapillary DGSS | average mIoU: 63.76 -> 63.32 | Hurt |
| use LoCA instead of full fine-tuning | MambaVision-B · GTAV to Cityscapes/BDD/Mapillary DGSS | average mIoU: 32.52 -> 44.20 | Improved |
| use LoCA instead of full fine-tuning | MobileMamba · ImageNet Top-1 accuracy | top-1 accuracy: 70.6% -> 78.9% | Improved |
| use LoCA r16 instead of r4 (ablation) | LoCA on ConvNeXt-B · VTAB-1k (19 tasks) | average accuracy: 75.98 -> 75.93 | No clear effect |
Closest papers by meaning
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.89)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.89)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.88)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.88)
- GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning (NeurIPS 2025 · similarity 0.87)
- Efficient Adaptation of Pre-trained Vision Transformer via Householder Transformation (NeurIPS 2024 · similarity 0.87)
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.
