SCOPE Framework for Complex Image Generation

A novel framework named SCOPE (Structured Decomposition and Conditional Skill Orchestration) tackles the issue of accurately translating intricate visual intents in text-to-image generation. The researchers highlight a "Conceptual Rift," where the semantic commitments—essential requirements that need to be monitored throughout grounding, generation, and verification—often become untraceable during the generation process. SCOPE addresses this by preserving these commitments within a dynamic structured specification and selectively utilizing retrieval, reasoning, and repair skills when commitments are either unresolved or breached. To assess the realization of commitment-level intent, the study presents Gen-Arena, a benchmark annotated by humans featuring entity- and constraint-level specifications. This research is available on arXiv under the identifier 2605.08043.