SSR-Merge

Give SSR-Merge a few LoRAs trained on the same base model and one text prompt per LoRA — no ground-truth images, no extra training. It runs a short calibration pass on one GPU and writes out a regular LoRA file you can load like any other.

Paper: SSR-Merge: Subspace Signal Routing for Training-Free LoRA Merging in Diffusion Models (ICML 2026). Zhengxuan Wei, Yi Dong, Zonghui Li, Xianhui Lin, Xing Liu, Hong Gu, Shaofeng Zhang, Wenbin Li, Qi Fan.

---

Installation

git clone https://github.com/nagara214/SSR-Merge.git
cd SSR-Merge
pip install -r requirements.txt

Requirements: Python 3.9+, PyTorch 2.1+, and diffsynth-studio.

---

Quick start

# Reproduce the paper's cat + dog merge on FLUX.1-dev (default):
python demo.py

# Or pick another backbone and bring your own LoRAs:
python demo.py --backbone qwen \
    --loras task_a.safetensors task_b.safetensors \
    --prompts "a photo of ..." "a photo of ..."

Before the default run, download the paper's cat and dog DreamBooth LoRAs from Google Drive and place the two .safetensors files in demo_loras/. The command then merges them into merged_flux.safetensors.

demo.py covers five DiT-based backbones out of the box: flux, qwen, z_image, hidream, flux2. The --loras arguments accept local paths or HuggingFace org/repo[:filename] specs.

A cat LoRA and a dog LoRA, merged by SSR into one LoRA that composes both subjects faithfully in a single image:

---

Citation

@inproceedings{wei2026ssrmerge,
  title={SSR-Merge: Subspace Signal Routing for Training-Free LoRA Merging in Diffusion Models},
  author={Wei, Zhengxuan and Dong, Yi and Li, Zonghui and Lin, Xianhui and Liu, Xing and Gu, Hong and Zhang, Shaofeng and Li, Wenbin and Fan, Qi},
  booktitle={ICML},
  year={2026},
}

License

Apache 2.0. See LICENSE.