SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation

SCOPE. Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation. By Yufei Jiao and seven others, from Xiamen University, the University of Rochester, and two other institutions. Published on the fifteenth of October. Its code repository is worked out: about one hundred and ninety source files, with a readme.

Reinforcement learning lets a language model improve text-to-image generation without touching the image generator at all: the model learns to rewrite prompts for a generator that stays frozen. Existing methods take a single signal derived from the resulting image and apply it to individual subword tokens. Yet the content of a prompt is not organised by subword. Objects, counts, attributes, and relations each form a coherent semantic span. Credit is being handed out at a finer grain than the meaning it is meant to reward.

SCOPE is an objective built on Group Relative Policy Optimization that aligns how updates are weighted with that structure. The rewriter generates two things together: an enhanced prompt, and a semantic partition of that prompt into spans.

Within each span, SCOPE pools the shifts in policy log probabilities to derive one shared coefficient for that span's token losses. Each completion still keeps a single advantage, so the reward signal itself is unchanged. It requires no separate reward for each span, and no additional image renders. What changes is only how the existing signal is spread across tokens.

The rewriter was tested with three generators: Qwen-Image, SD3, and FLUX.1-dev. Averaged across them, SCOPE improves on vanilla GRPO by a relative nine point four eight percent on GeneVal and seven point six two percent on T2I-CompBench. Ablation studies show that both pieces matter: the semantic boundaries, and the pooling within spans, each contribute to those gains.

The change is narrow on purpose. Reward stays as it was, rendering stays as it was, and only the weighting of token updates now follows units of meaning rather than subword pieces. The authors present semantic structure as a practical inductive bias for prompt engineering with reinforcement learning, one that costs nothing extra at the image generator.