SCOPE: GRPO-based prompt rewriting weighted by semantic spans for text-to-image
An arXiv paper submitted on 15 October 2026, "SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation" by Yufei Jiao, Jinfa Huang and six co-authors, targets reinforcement-learning prompt rewriting for a frozen text-to-image generator. The authors argue that existing methods apply a single image-derived signal to individual subword tokens, even though objects, counts, attributes and relations form coherent semantic spans. SCOPE builds on Group Relative Policy Optimization: the rewriter jointly produces an enhanced prompt and its semantic partition, pools shifts in policy log probabilities within each span to derive a shared coefficient for its token losses, and keeps a single advantage per completion, without separate span rewards or additional image renders. The authors report average relative gains over vanilla GRPO of 9.48% on GeneVal and 7.62% on T2I-CompBench across Qwen-Image, SD3 and FLUX.1-dev, with ablations attributing the gains to both semantic boundaries and within-span pooling.
The authors report that span-level pooling of update weights improved results over vanilla GRPO across three image generators, which they present as evidence that semantic structure is a practical inductive bias for RL-based prompt engineering.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.22264 (cs)
[Submitted on 15 Oct 2026]
Title:SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation
Authors:Yufei Jiao, Jinfa Huang, Zeyue Xue, Tianyu Xie, Yuexiao Ma, Rongrong Ji, Xiawu Zheng, Jiebo Luo
View a PDF of the paper titled SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation, by Yufei Jiao and 7 other authors