Inquiry about GPT-4.1 Rule Prompt for Image Editing GRPO Stage

[SuJieHu]

Hello,
First of all, congratulations on the excellent work with UniPic2 and Skywork-EditReward — it’s truly inspiring research.
I was particularly impressed by the reinforcement learning stage for image editing, where you mention employing GPT-4.1 with carefully designed evaluation templates to provide multi-dimensional scoring (instruction-following accuracy, image quality, etc.) before training Skywork-EditReward.

I would like to kindly ask:

1. During the GRPO (reinforcement learning) stage for image editing, did you use GPT-4.1 directly as the rule-based evaluator to provide reward signals?

2. If possible, could you share (or point me to) the evaluation template / rule prompt you used with GPT-4.1 for scoring the generated images?

Having access to the prompt design would greatly help me and others in the community better understand and potentially reproduce the evaluation pipeline you proposed.
Thank you very much for your time and for making this valuable work available.

Best regards,

[nonwhy]

Hello, First of all, congratulations on the excellent work with UniPic2 and Skywork-EditReward — it’s truly inspiring research. I was particularly impressed by the reinforcement learning stage for image editing, where you mention employing GPT-4.1 with carefully designed evaluation templates to provide multi-dimensional scoring (instruction-following accuracy, image quality, etc.) before training Skywork-EditReward.

I would like to kindly ask:

1. During the GRPO (reinforcement learning) stage for image editing, did you use GPT-4.1 directly as the rule-based evaluator to provide reward signals?
2. If possible, could you share (or point me to) the evaluation template / rule prompt you used with GPT-4.1 for scoring the generated images?

Having access to the prompt design would greatly help me and others in the community better understand and potentially reproduce the evaluation pipeline you proposed. Thank you very much for your time and for making this valuable work available.

Best regards,

Thank you so much for your interest in UniPic2! Here are the answers to your questions:

1. We experimented with directly using GPT-4.1 for online evaluation to provide reward signals for image editing tasks. Please refer to the ablation study section in our paper for details.

2. Regarding the evaluation template used for GPT-4.1 scoring, as the text is too lengthy to display here, we will soon update it in the GitHub code repository. You can refer to the Step1X-Edit/GEdit-Bench for reference.