Research map / Image generation / Editing and style transfer
Spanning the Visual Analogy Space with a Weight Basis of LoRAs
Hila Manor, Rinon Gal, Haggai Maron, Tomer Michaeli, Gal Chechik
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 |
|---|---|---|---|
| use basis of 32 rank-4 LoRAs instead of one r=128 LoRA | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.70 -> 5.94 | Improved |
| use basis of 32 rank-4 LoRAs instead of one r=128 LoRA | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Pairwise VLM vs LoRA r=128: 57.9 | Improved |
| add larger basis N=64 (ablation) | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.94 -> 5.48 | Hurt |
| add larger rank r=16 (N=32 equivalent) (ablation) | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.94 -> 4.92 | Hurt |
| use large basis N=32 instead of N=8 at same capacity (ablation) | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.49 -> 5.94 | Improved |
| use N=32 instead of N=16 (ablation) | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.95 -> 5.94 | No clear effect |
| use softmax instead of Tanh for mixing coefficients (ablation) | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 4.49 -> 5.94 | Improved |
| encode a, a', b separately and concatenate instead of encoding the 2x2 grid (ablation) | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.75 -> 5.94 | Improved |
| swap CLIP for SigLIP2 encoder (ablation) | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.94 -> 5.82 | No clear effect |
| add mixture of LoRAs vs single LoRA r=256 | LoRWeB vs LoRA · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 5.48 -> 5.94 | Improved |
| use LoRWeB instead of DIA-Kontext | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | Edit accuracy (VLM): 3.56 -> 5.94 | Improved |
| use LoRWeB instead of PairEdit s=0.6 | LoRWeB · Visual analogy eval (100 tasks, 840 triplets) | LPIPS: 0.24 vs 0.31 | Hurt |
Closest papers by meaning
- RelationAdapter: Learning and Transferring Visual Relation with Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
- LoRA Recycle: Unlocking Tuning-Free Few-Shot Adaptability in Visual Foundation Models by Recycling Pre-Tuned LoRAs (CVPR 2025 · similarity 0.87)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.87)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.86)
- LoRASculpt: Sculpting LoRA for Harmonizing General and Specialized Knowledge in Multimodal Large Language Models (CVPR 2025 · similarity 0.86)
- Text-to-LoRA: Instant Transformer Adaption (ICML 2025 · similarity 0.86)
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.
