Research map / Label-efficient and robust learning / Domain adaptation and generalization
Concept-Aware LoRA for Domain-Aligned Segmentation Dataset Generation
Minho Park, Sunghyun Park, Jungsoo Lee, Hyojin Park, Kyuwoong Hwang, Fatih Porikli, Jaegul Choo, Sungha Choi
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 |
|---|---|---|---|
| select layers by concept awareness (top 2%) instead of LoRA on all projection layers (ablation) | CA-LoRA · Cityscapes 0.3% in-domain | mIoU: 42.97 -> 44.13 | Improved |
| fine-tune a random 2% subset of layers instead of concept-aware selection (ablation) | CA-LoRA · Cityscapes 0.3% in-domain | mIoU: 44.13 -> 43.24 | Hurt |
| add diffusion-gradient normalization to concept awareness (ablation) | CA-LoRA · Cityscapes 0.3% in-domain | mIoU: 43.01 -> 44.13 | Improved |
| fine-tune only Q projections (hand-crafted) (ablation) | DatasetDM with LoRA · Cityscapes 0.3% in-domain | mIoU: 42.82 -> 40.50 | Hurt |
| fine-tune only K projections (hand-crafted) (ablation) | DatasetDM with LoRA · Cityscapes 0.3% in-domain | mIoU: 42.82 -> 43.50 | Improved |
| measure concept awareness at timestep 81 instead of 1, 201 or 481 (ablation) | CA-LoRA · Cityscapes 0.3% in-domain | mIoU: 41.68 (t=201), 40.80 (t=481), 40.49 (t=1) -> 44.13 (t=81) | Improved |
| increase the proportion of fine-tuned layers from 2% to 10% or 100% (ablation) | CA-LoRA · Cityscapes 0.3% in-domain | mIoU: 44.13 -> 43.36 (10%), 42.97 (100%) | Hurt |
| use projection-wise grouping instead of block-wise, layer-wise or integrated projection-wise (ablation) | CA-LoRA · Cityscapes 0.3% in-domain | mIoU: 43.67 (block-wise) -> 44.13 | Improved |
| use Viewpoint CA-LoRA rather than Style CA-LoRA for domain generalization (ablation) | CA-LoRA · adverse-weather text adherence | CLIP-Score | Improved |
| fine-tune the T2I model with CA-LoRA before training the label generator | DatasetDM · Cityscapes in-domain | mIoU: 42.82 -> 44.13 (0.3%) | Improved |
| standard LoRA and AdaLoRA fine-tuning of the T2I model | DatasetDM · Cityscapes in-domain | mIoU: AdaLoRA 3%: +1.86; 100%: -0.78 | Mixed |
| InstructPix2Pix generated data | Mask2Former · Cityscapes in-domain | mIoU: 10%: -2.81; 3%: +1.36 | Mixed |
Closest papers by meaning
- Semantic Library Adaptation: LoRA Retrieval and Fusion for Open-Vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.86)
- Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model (ICLR 2024 · similarity 0.86)
- DynAlign: Unsupervised Dynamic Taxonomy Alignment for Cross-Domain Segmentation (ICLR 2025 · similarity 0.86)
- Discriminative Probing and Tuning for Text-to-Image Generation (CVPR 2024 · similarity 0.85)
- Cached Multi-Lora Composition for Multi-Concept Image Generation (ICLR 2025 · similarity 0.85)
- A Generative Approach at the Instance-Level for Image Segmentation Under Limited Training Data Conditions (Student Abstract) (AAAI 2025 · similarity 0.85)
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.
