MT-OPSD checkpoints

LoRA checkpoints for MT-OPSD: On-Policy Self-Distillation for Multi-Turn Image Editing. MT-OPSD trains an editor on its own self-generated multi-turn states, which keeps it following instructions and keeps its images intact over long editing sessions.

Subfolder	Base model	LME-Bench SR@10 (base → MT-OPSD)	CR@10 (base → MT-OPSD)
qwen-image-edit-2511/	Qwen/Qwen-Image-Edit-2511	0.03 → 0.44	0.55 → 0.02
firered-image-edit-1.0/	FireRedTeam/FireRed-Image-Edit-1.0	0.15 → 0.52	0.61 → 0.03
flux2-klein-base-9b/	black-forest-labs/FLUX.2-klein-base-9B	0.12 → 0.38	0.25 → 0.04

Each subfolder holds pytorch_lora_weights.safetensors (rank 32, alpha 64, fp32, diffusers format) and adapter_config.json (target modules and the sampling settings used in training). Use of each LoRA is also subject to the license of its base model.

Usage

git clone https://github.com/liangbingzhao/MT-OPSD.git && cd MT-OPSD
pip install -r requirements.txt
hf download metazlb/MT-OPSD --include "qwen-image-edit-2511/*" --local-dir checkpoints

python inference.py --ckpt checkpoints/qwen-image-edit-2511 --image input.png --output_dir outputs/demo \
    --instructions "Make it a snowy winter scene." "Add a red scarf to the dog." "Convert to a pencil sketch."

Use the settings the checkpoints were trained with (inference.py reads them from adapter_config.json): a 512×512 working area (aspect preserved), 30 sampling steps and true CFG 4.0. Keep the LoRA in fp32 with bf16 autocast, as inference.py does; casting the LoRA to bf16 measurably weakens long-horizon robustness.

Citation

@article{zhao2026mtopsd,
  title={MT-OPSD: On-Policy Self-Distillation for Multi-Turn Image Editing},
  author={Zhao, Liangbing and Zhuo, Le and Elhoseiny, Mohamed},
  year={2026}
}