Research map / Architectures and efficiency / Pruning and efficient tuning
SSR-Merge: Subspace Signal Routing for Training-Free LoRA Merging in Diffusion Models
Zhengxuan Wei, Yi Dong, Zonghui Li, Xianhui Lin, Xing Liu, Hong Gu, Shaofeng Zhang, Wenbin Li, Qi Fan
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 identity router (Task Arithmetic) with statistics-derived router (ablation) | SSR on FLUX.1-dev · DreamBooth 10 subjects, K=9 merged | DINOv2 similarity: 0.5356 -> 0.6713 | Improved |
| use SSR instead of DARE | FLUX.1-dev LoRA merging · DreamBooth 10 subjects, K=9 merged | CLIP score: 0.7376 -> 0.7850 | Improved |
| increase calibration steps from 1 to 20 (ablation) | SSR K=9 · DreamBooth 10 subjects, K=9 merged, FLUX.1-dev | DINOv2 similarity: 0.6713 -> 0.6739 | No clear effect |
| use more calibration timesteps (20 vs 1) (ablation) | SSR K=9 · FLUX.1-dev K=9 calibration | merge time: >600s vs 34.26 seconds | Hurt |
| increase merged LoRAs from 9 to 21 | SSR · DreamBooth 10 subjects, FLUX.1-dev | DINOv2 recovery rate: 90.2% -> 77.0% | Hurt |
| use SSR instead of DARE in multi-concept composition | FLUX.1-dev · Multi-concept composition (K=2-4) | Success rate: 0.62 -> 0.91 | Improved |
| use SSR instead of DARE for facial editing | FLUX.1-dev · FFHQ facial editing | ArcFace similarity: 0.9471 -> 0.9610 | Improved |
| vary calibration prompt template (five templates) (ablation) | SSR K=3 · DreamBooth, K=3 | CLIP score: 0.7915 to 0.8267, var 0.00067 | No clear effect |
| add Tikhonov regularization lambda<=0.1 to RegMean | RegMean on FLUX.1 · DreamBooth merge | image quality | Hurt |
| use SSR instead of TIES on Qwen-Image | Qwen-Image · DreamBooth 10 subjects, K=9 merged | DINOv2 similarity: 0.5716 -> 0.7461 | Improved |
| use SSR instead of DOGE TA on GLUE | 8 GLUE tasks · GLUE (8 tasks) | average score: 79.9 -> 80.9 | Improved |
| use SSR instead of TIES for merge time | FLUX.1-dev K=9 · merge wall-clock | seconds: 88.93 -> 34.26 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- IterIS — IterIS: Iterative Inference-Solving Alignment for LoRA Merging (CVPR 2025) · better on 4 of 4 shared results
- RobustMerge — RobustMerge: Parameter-Efficient Model Merging for MLLMs with Direction Robustness (NeurIPS 2025) · better on 2 of 2 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
- LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · similarity 0.87)
- Crowded in B-Space: Calibrating Shared Directions for LoRA Merging (NeurIPS 2026 · similarity 0.86)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.86)
- IterIS: Iterative Inference-Solving Alignment for LoRA Merging (CVPR 2025 · similarity 0.85)
- Accurate and Efficient Low-Rank Model Merging in Core Space (NeurIPS 2025 · similarity 0.85)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 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.
