Research map / Image generation / Text-to-image diffusion
Contrastive Test-Time Composition of Multiple LoRA Models for Image Generation
Tuna Han Salih Meral, Enis Simsar, Federico Tombari, Pinar Yanardag
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 |
|---|---|---|---|
| CLoRA vs strongest baseline Mix-of-Show | Stable Diffusion v1.5 multi-LoRA · 131 LoRAs, 200 two-subject prompts | Avg DINO similarity: 0.526 -> 0.554 | Improved |
| CLoRA vs strongest baseline Mix-of-Show | Stable Diffusion v1.5 multi-LoRA · 131 LoRAs, 200 two-subject prompts | Avg CLIP-I: 0.712 -> 0.725 | Improved |
| CLoRA vs Composite on text alignment | Stable Diffusion v1.5 multi-LoRA · 131 LoRAs, 200 two-subject prompts | CLIP-T: 0.833 -> 0.862 | Improved |
| CLoRA vs ZipLoRA in user study | Stable Diffusion v1.5 multi-LoRA · 131 LoRAs, 200 two-subject prompts | User study faithfulness (1-5): 2.81 -> 3.32 | Improved |
| CLoRA vs Mix-of-Show on segmented instances | Stable Diffusion v1.5 multi-LoRA · SEEM instance-segmented compositions | Avg CLIP similarity (%): 78.7% -> 87.1% | Improved |
| CLoRA vs Mix-of-Show and Orthogonal Adaptation with 3-5 LoRAs | Stable Diffusion v1.5 multi-LoRA · 3-5 LoRA compositions | Avg DINO similarity: 0.351 -> 0.347 | No clear effect |
| increase LoRAs from 2 to 8 | CLoRA · scalability | runtime / VRAM: 24s -> 96s; 25GB -> 80GB | Hurt |
| add Latent Update (contrastive objective) (ablation) | CLoRA · qualitative | composition correctness | Improved |
| add Latent Masking (ablation) | CLoRA · qualitative | subject identity | 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 6 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
- Cached Multi-Lora Composition for Multi-Concept Image Generation (ICLR 2025 · similarity 0.91)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.91)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.89)
- LoRA Recycle: Unlocking Tuning-Free Few-Shot Adaptability in Visual Foundation Models by Recycling Pre-Tuned LoRAs (CVPR 2025 · similarity 0.89)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.89)
- LoRACLR: Contrastive Adaptation for Customization of Diffusion Models (CVPR 2025 · similarity 0.88)
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.
