Research map / Image generation / Text-to-image diffusion
LoRA.rar: Learning to Merge LoRAs via Hypernetworks for Subject-Style Conditioned Image Generation
Donald Shenaj, Ondrej Bohdal, Mete Ozay, Pietro Zanuttigh, Umberto Michieli
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 test-time optimization (ZipLoRA) with a pre-trained hypernetwork predicting merging coefficients | LoRA merging · MARS2 test pairs (SDXL) | MARS2 average case: 0.58 -> 0.71 | Improved |
| replace per-pair optimization with a single forward pass of the hypernetwork | LoRA merging · Footprint | time to predict merging coefficients: 158s -> 0.037s | Improved |
| apply LoRA.rar on a lightweight backbone | LoRA.rar · KOALA 700m | MARS2 average case: 0.51 -> 0.56 | Improved |
| hypernetwork-merge query and output LoRAs, average key and value (ablation) | LoRA.rar · Validation pairs | MARS2 average case: 0.23 -> 0.48 (all four vs Query, Output) | Improved |
| use hypernetwork on query, value and output only (ablation) | LoRA.rar · Validation pairs | MARS2 best case: 0.92 -> 0.33 | Hurt |
| use MARS2 instead of CLIP-I or DINO to track human preference | Evaluation metric · Direct merge human study | correlation with human score: CLIP-I 0.08, DINO -0.01, MARS2 0.76 | Improved |
| train hypernetwork on narrow splits with no category overlap (ablation) | LoRA.rar · New splits | MARS2 | Mixed |
| LoRA.rar on unseen furniture subjects and cyberpunk style | LoRA.rar · New concepts | MARS2: 0.6 -> 0.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ZipLoRA — ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs (ECCV 2024) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs (CVPR 2025 · similarity 0.89)
- NP-LoRA: Null Space Projection for Subject-Style LoRA Fusion (NeurIPS 2026 · similarity 0.88)
- LoRACLR: Contrastive Adaptation for Customization of Diffusion Models (CVPR 2025 · similarity 0.88)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.88)
- Text-to-LoRA: Instant Transformer Adaption (ICML 2025 · similarity 0.88)
- Cached Multi-Lora Composition for Multi-Concept Image Generation (ICLR 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.
