SCOPE: Semantic-Unit Credit Optimization for Prompt Rewriting in Text-to-Image Generation

Problem

When a prompt rewriting model is trained with reinforcement learning, the single reward from one image is applied equally to every subword token. But objects, counts, attributes and relations are semantic units that span multiple tokens. Because token boundaries and semantic boundaries do not match, credit gets fragmented. On the other hand, assigning a separate reward to each span requires rendering additional images, which is costly.

Method

The rewriting model (Qwen3-8B) was made to output the improved prompt as semantic spans separated by '|'. A separate format reward is given for the output format. Within each span, the change in log-probability between the current and previous policy is averaged to produce a coefficient. This coefficient is clipped to [0.8, 1.2], normalized so that its token-weighted average equals 1, and then detached from the gradient. A single completion-level advantage is kept, and the coefficient only adjusts how much the tokens within a span contribute to the loss. No per-span rewards, extra rendering or learned critic are needed. With sensitivity β=0, the method reduces to standard GRPO. The image generation models (Qwen-Image, SD3, FLUX.1-dev) were kept frozen.

Results

• Compared with standard GRPO, the overall GenEval score rose from 0.91→0.95 for Qwen-Image, 0.74→0.84 for SD3 and 0.76→0.84 for FLUX.1-dev. The average relative improvement across the three models was 9.48%.
• On T2I-CompBench, all three models scored higher than GRPO in all 6 categories, with an average relative improvement of 7.62%.
• On SD3, counting accuracy rose sharply from 0.58→0.79 and position accuracy from 0.57→0.83. The method beat GRPO on 17 of the 18 GenEval sub-items.
• In ablations, grouping at the span level worked better than per-token weights, and semantic boundaries worked better than random or fixed-length boundaries. The authors presented this as evidence that the semantic structure itself drives the effect.

Limitations

Because the GenEval and T2I-CompBench evaluation models were used directly as training rewards, it is hard to rule out that the gains are tuned to the benchmark evaluators. The prompt sets themselves are kept separate between training and evaluation. Only one rewriting model, Qwen3-8B, was tested. In the RePrompt comparison, the SD3 and FLUX.1-dev numbers were taken from the published paper rather than reproduced. Image quality and human preference evaluations could not be found in the provided text.

Why it matters

It is a low-cost way to improve count, position and attribute alignment by improving only the prompt layer, without retraining the image generation model. Since it can be added to existing GRPO pipelines without extra rendering, the barrier to practical adoption is low.

Key terms

GRPO(Group Relative Policy Optimization)

A reinforcement learning technique that generates several responses for one input and computes the advantage by comparing rewards within that group. It does not need a separate value model (critic).

Credit Assignment

The problem of dividing a single final reward according to how much each part of a response (tokens or segments) contributed to it.

Semantic span

A phrase-level unit that forms a single meaning, such as an object, count, attribute or spatial relation. It can span multiple subword tokens.

Prompt Rewriter

A language model that rewrites a user's prompt into a form the image generation model follows better. It is trained while the image generation model is kept frozen.

Original

• arXiv abstract
• PDF
• HTML full text
• Hugging Face

Authors: Yufei Jiao, Jinfa Huang, Zeyue Xue, Tianyu Xie, Yuexiao Ma, Rongrong Ji, Xiawu Zheng, Jiebo Luo · Posted 2026-10-15 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.