SCOPE: semantic span-aware GRPO objective improves T2I prompt rewriting
university-of-rochester · hf · 2026-10-16
Researchers from Xiamen University, University of Rochester and others propose SCOPE, a semantic bounded credit optimization method for prompt engineering in text-to-image generation.
- Instead of assigning a single image-derived signal to individual subword tokens, SCOPE aligns GRPO policy updates with coherent semantic spans
- The rewriter jointly generates the enhanced prompt and its semantic partition; shifts in log probabilities within each span are pooled into a shared coefficient
- Improves generation quality without extra rendering
More from Multimodal
- Alibaba open-sources CommerceGen-Studio for commercial-grade e-commerce images — alibabagroup · 2026-10-16
- EVA-Cap: Fudan and OPPO align video captioning with atomic audiovisual events via GRPO — Fudan-University · 2026-10-16
- SCOPED: skill-decoupled expert distillation beats mixed-capability RL across 8 benchmarks — Yuqi Xiong · 2026-10-16
- ComfyUI v0.41.0 ships day-one FLUX.3 Image support with sampler presets — lmoroney · 2026-10-16
- Krea2 Realism returns as a full checkpoint with natural skin textures and usage tips — rynaleopard · 2026-10-16
- Removing context conditioning makes image conditioning stronger in flow-guided video generation — pixlpa · 2026-10-16