Research map / Vision-language and multimodal / CLIP and zero-shot
DeFT-LoRA: Decoupled and Fused Tuning with LoRA Experts for Universal Cross-Domain Retrieval
Ke Xu, Xiaozheng Shen, Shanshan Wang, Mengzhu Wang, Xun Yang
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 |
|---|---|---|---|
| replace single LoRA with decoupled base+domain experts (ablation) | DeFT-LoRA · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.5909 -> 0.6093 | Improved |
| replace naive LoRA-MoE with three-stage base+domain experts (ablation) | DeFT-LoRA · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.5970 -> 0.6093 | Improved |
| add Base Expert (ablation) | DeFT-LoRA · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.6014 -> 0.6093 | Improved |
| use dynamic gating instead of static averaging (ablation) | DeFT-LoRA · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.6019 -> 0.6093 | Improved |
| share A matrix across experts instead of independent A (ablation) | DeFT-LoRA · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.6070 -> 0.6093 | Improved |
| apply LoRA to deep layers 7-12 instead of shallow 1-6 (ablation) | DeFT-LoRA · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.5558 -> 0.6068 | Improved |
| apply LoRA to all layers instead of only deep layers (ablation) | DeFT-LoRA · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.6068 -> 0.6093 | Improved |
| add DeFT-LoRA to zero-shot CLIP | CLIP ViT-B/32 · DomainNet UCDR (Infograph unseen, Unseen Gallery) | mAP@200: 0.5007 -> 0.6093 | Improved |
| DeFT-LoRA vs UCDR-Adapter | DeFT-LoRA · DomainNet UCDR (Painting unseen, Mixed Gallery) | mAP@200: 0.7203 vs 0.7135 | Hurt |
| DeFT-LoRA vs ProS/UCDR-Adapter | DeFT-LoRA · DomainNet UCDR (Sketch unseen, Mixed Gallery) | mAP@200: 0.6073 vs 0.6014 | Hurt |
| DeFT-LoRA vs UCDR-Adapter | DeFT-LoRA · DomainNet UCDR average (Unseen Gallery) | mAP@200: 0.6071 -> 0.6212 | Improved |
| use LoRA experts instead of prompts (efficiency) | DeFT-LoRA · DomainNet Infograph | trainable parameters (M): 2.36 -> 1.29 | Improved |
Closest papers by meaning
- DA-MergeLoRA: Hypernetwork-Based LoRA Merging for Few-Shot Test-Time Domain Adaptation (ECCV 2026 · similarity 0.88)
- Retrieval Across Any Domains via Large-scale Pre-trained Model (ICML 2024 · similarity 0.88)
- Unsupervised Feature Representation Learning for Domain-generalized Cross-domain Image Retrieval (ICCV 2023 · similarity 0.88)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.87)
- Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.87)
- Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation (AAAI 2025 · 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.
