README.md

metadata

license: apache-2.0
library_name: diffusers
pipeline_tag: text-to-image
tags:
  - lora
  - image-generation
  - reinforcement-learning
  - reward-modeling
  - firm

FIRM-SD3.5

This repository contains the LoRA weights for FIRM-SD3.5, an enhanced text-to-image generation model developed using the FIRM (Faithful Image Reward Modeling) framework.

The model was introduced in the paper Trust Your Critic: Robust Reward Modeling and Reinforcement Learning for Faithful Image Editing and Generation.

• Project Page: https://firm-reward.github.io/
• GitHub Repository: https://github.com/VisionXLab/FIRM-Reward
• Paper: arXiv:2603.12247

Model Description

Reinforcement learning (RL) for visual generation relies heavily on the faithfulness of the reward model used as a critic. FIRM addresses common issues like hallucinations and noisy scoring through:

1. Tailored Data Pipelines: Using specialized curation for editing (execution and consistency) and generation (instruction following).
2. Robust Reward Models: Training specialized reward models (like FIRM-Gen-8B) on high-quality scoring datasets.
3. "Base-and-Bonus" Reward Strategy: A novel strategy to balance competing objectives, such as Quality-Modulated Alignment (QMA) for generation.

The resulting FIRM-SD3.5 model demonstrates significant breakthroughs in fidelity and instruction adherence compared to existing general models by mitigating hallucinations.

Citation

@article{zhao2026trust,
  title={Trust Your Critic: Robust Reward Modeling and Reinforcement Learning for Faithful Image Editing and Generation},
  author={Zhao, Xiangyu and Zhang, Peiyuan and Lin, Junming and Liang, Tianhao and Duan, Yuchen and Ding, Shengyuan and Tian, Changyao and Zang, Yuhang and Yan, Junchi and Yang, Xue},
  journal={arXiv preprint arXiv:2603.12247},
  year={2026}
}

Acknowledgements

This project was developed by the VisionXLab and builds upon several open-source projects including flow-grpo, DiffusionNFT, and Edit-R1.