QWW/DivRL

Model Card for Model ID

This is the pre-trained model weight for paper DivRL: Disentangled Self-Similarity Rewards for Diverse Subject-Driven Generation.

• Finetuned from model: FLUX.1-Kontext-dev
• Repository: https://github.com/QianWangX/DivRL
• Paper: https://arxiv.org/abs/2606.23950
• Demo: https://qianwangx.github.io/DivRL/

Uses

Stage-1 weight is trained with nSSM as reward model only. Built on top of Stage-1 weight, Stage-2 weight is further trained on nSSM + VSM collaboratively to obtain the final results shown in the paper.

You can refer to the Stage-1 weight for generation with high diversity but low consistency, and the Stage-2 weight for generation with both high diversity and high consistency.

How to Get Started with the Model

Please refer to https://github.com/QianWangX/DivRL.

Training Details

Training Data

We provide the training data at QWW/Syncd_filtered.

BibTeX:

@misc{wang2026divrl,
      title={DivRL: Disentangled Self-Similarity Rewards for Diverse Subject-Driven Generation}, 
      author={Qian Wang and Zhenyu Li and Abdelrahman Eldesokey and Peter Wonka},
      year={2026},
      eprint={2606.23950},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.23950}, 
}