Original title: GDB-Reward: From Evaluation Metrics to Training Rewards for Graphic Design
Adrienne Deganutti, Purvanshi Mehta, Simon Hadfield, Andrew Gilbert · September 2026
Standard text-to-image AI tools create great photos, but they struggle with graphic design. Good graphic design requires meeting exact rules for typography, layout, color, and visual communication. Changing the base image generator directly is expensive. Instead, researchers wanted to teach the system better prompts without changing the image generator at all. The problem is that creating images this way does not allow standard step-by-step mathematical tuning. To solve this, the researchers created GDB-Reward. This system takes separate, regular graphic design scoring rules and turns them into single scores for reinforcement learningA learning method where an AI improves by trying actions and receiving reward signals., a method where AI learns through trial and error. As a result, the frozen generator learned to follow design rules much better, boosting visual quality, rendering fidelity, and layout accuracy.
Imagine baking a cake where you follow a recipe but can only taste the final result. If the cake is too salty or lopsided, you cannot adjust the baking process mid-way. In this paper, the evaluation metrics are like the recipe standards, and reinforcement learning uses the final taste score to learn how to tweak the recipe for the next cake without changing the oven itself.
Graphic design demands exact rules for text, layout, and color, but AI often misses these, forcing manual fixes. This lets AI learn from design scores to follow specs better, saving time and cost for designers and businesses.
The abstract doesn't specify limitations, but it focuses on a framework for graphic design metrics. It doesn't address broader applications or potential biases in the evaluation metrics.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
GDB-Reward: From Evaluation Metrics to Training Rewards for Graphic Design — Adrienne Deganutti, Purvanshi Mehta, Simon Hadfield et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.