Back to Project
S

SCOPE

Project Oct 16, 2026

A research training objective that assigns semantic-span credit to a Qwen3-8B prompt rewriter for frozen text-to-image generators.

Visit primary link 3 sources attached
Authors' generated image comparison for a person with an apple, a cup above a couch, four handbags and a white orange, with original prompts, GRPO, gpt-image 2 and SCOPE outputs.
Author-generated selected examples from SCOPE, including the rewritten prompt and comparison labels. These images are research illustrations, not an AITrending reproduction or evidence of an overall win over gpt-image 2. Resized without cropping. Source: Yufei Jiao, Jinfa Huang and collaborators / SCOPE
multimodalself-hosted

What happened lately

SCOPE timeline

SCOPE, Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation, trains a language model to rewrite image prompts. It groups objects, counts, attributes and relations into semantic spans, then uses span weights when updating the rewriter. The image generator remains frozen.

The paper evaluates Qwen3-8B rewrites with Qwen-Image, Stable Diffusion 3 and FLUX.1-dev. Results and generated comparisons are author reported. Public training code documents Linux x86_64, NVIDIA GPUs, Python 3.10 and separate training, rendering and evaluation environments. The README does not link a ready-trained SCOPE rewriter checkpoint. A repository-wide reuse license was not established at review; licenses bundled with upstream benchmarks do not establish one for the entire project.