Large-Scale Universal Defect Generation: Foundation Models and Datasets
Paper โข 2604.08915 โข Published
Defect-Und-Reward checkpoint for UniDG (UniDG: Unified Defect Generation via Defect-Context Editing, NeurIPS 2026), fine-tuned from Qwen/Qwen3-VL-4B-Instruct with LLaMA-Factory on 13,786 curated defect-editing evaluation conversations.
[Paper] [Code] [UniDG-SFT-LoRA] [UniDG-RFT-LoRA]
During Consistency-RFT (Flow-GRPO), Defect-Und-Reward evaluates each synthesized defect image alongside the reference image, reference mask, and target mask across four dimensions (0โ5 scale):
Download this repository and set reward_model in train/rft/config.yaml:
hf download retofan23333/UniDG-Defect-Und-Reward-Release --local-dir checkpoints/UniDG-Defect-Und-Reward-Release
@inproceedings{fan2026unidg,
title={UniDG: Unified Defect Generation via Defect-Context Editing},
author={Fan, Yuanting and Liu, Jun and Gao, Bin-Bin and Chen, Xiaochen and Lin, Yuhuan and Dai, Zhewei and Zhan, Jiawei and Wang, Chengjie},
booktitle={NeurIPS},
volume={39},
year={2026},
publisher={Curran Associates, Inc.}
}
Base model
Qwen/Qwen3-VL-4B-Instruct