Source-linked AI summary
SCOPE: Structured Decomposition and Conditional Skill Orchestration for Complex Image Generation
Tianfei Ren, Zhipeng Yan, Yiming Zhao, Zhen Fang, Yu Zeng, Guohui Zhang, Hang Xu, Xiaoxiao Ma, Shiting Huang, Ke Xu, Wenxuan Huang, Lionel Z. Wang, Lin Chen, Zehui Chen, Jie Huang, Feng Zhao
TL;DR
Faithfully realizing complex visual intent requires preserving many commitments across grounding, generation, and verification, which existing evaluations and workflows do not consistently track. SCOPE maintains these commitments in an evolving specification and conditionally orchestrates retrieval, reasoning, repair, and verification, substantially outperforming evaluated baselines on Gen-Arena with 0.60 EGIP and achieving strong external-benchmark results.
Problem
Complex image generation must preserve defining entities and constraints, but coarse holistic scores and largely independent checklists provide limited commitment-level evaluation.
Method
SCOPE maintains semantic commitments in an evolving structured specification and conditionally orchestrates retrieval, reasoning, repair, and verification around unresolved or violated commitments.
Results
SCOPE substantially outperforms all evaluated baselines on Gen-Arena, achieving 0.60 EGIP, and also performs strongly on WISE-V and MindBench.
Takeaways & Limitations
Keeping commitments identifiable across the generation lifecycle is a practical path toward more faithful complex image generation.
Takeaways & Limitations
SCOPE incurs higher inference-time cost and latency than one-shot generation and depends on verifier accuracy, whose errors can misroute repair actions.
Abstract
from arXiv · showhide
While text-to-image models have made strong progress in visual fidelity, faithfully realizing complex visual intents remains challenging because many requirements must be tracked across grounding, generation, and verification. We refer to these requirements as semantic commitments and formalize their lifecycle discontinuity as the Conceptual Rift, where commitments may be locally resolved or checked but fail to remain identifiable as the same operational units throughout the generation lifecycle. To address this, we propose SCOPE, a specification-guided skill orchestration framework that maintains semantic commitments in an evolving structured specification and conditionally invokes retrieval, reasoning, and repair skills around unresolved or violated commitments. To evaluate commitment-level intent realization, we introduce Gen-Arena, a human-annotated benchmark with entity- and constraint-level specifications, together with Entity-Gated Intent Pass Rate (EGIP), a strict entity-first pass criterion. SCOPE substantially outperforms all evaluated baselines on Gen-Arena, achieving 0.60 EGIP, and further achieves strong results on WISE-V (0.907) and MindBench (0.61), demonstrating the effectiveness of persistent commitment tracking for complex image generation.
1 Introduction
The introduction frames complex image generation as a semantic-commitment continuity problem: requirements may emerge across grounding, reasoning, realization, and checking but lose identity across the lifecycle. SCOPE addresses this Conceptual Rift with persistent structured specifications and conditional skills, while Gen-Arena evaluates commitment-level realization using entity-first criteria.
- Motivation: Complex visual intents require semantic commitments that become actionable at different stages and remain identifiable until visually realized and checked.Some requirements are explicit initially, while others emerge through external grounding or scene reasoning.
- SCOPE: SCOPE maintains semantic commitments in an evolving structured specification and conditionally orchestrates retrieval, reasoning, repair, and verification skills.These skills ground missing external information, infer implicit requirements, revise violated commitments, and verify generated results.
- Evaluation: Gen-Arena is a human-annotated benchmark with entity- and constraint-level specifications and Entity-Gated Intent Pass Rate (EGIP) as a strict entity-first criterion.Linking constraints to prerequisite entities enables diagnosis of missing entities, violated constraints, and downstream failures.
- Problem: The Conceptual Rift describes how semantic commitments lose continuity across the complex image-generation lifecycle.SCOPE targets this rift by preserving commitments as operational units across generation stages.
- Results: 0.60 EGIP is achieved by SCOPE on Gen-Arena, alongside WISE-V (0.907) and MindBench (0.61), outperforming all evaluated baselines on Gen-Arena.The reported results support persistent commitment tracking for complex image generation.
2 Related Work
Recent multimodal agents mediate complex image generation through request interpretation, retrieval-grounded reasoning, and other interventions. However, existing intermediate representations are typically intervention-specific, limiting persistent tracking of semantic commitments across generation, while reusable skill abstractions support broader visual workflows.
- Multimodal agents: Multimodal agents increasingly mediate complex image generation beyond direct prompt conditioning through request interpretation, retrieval grounding, and multimodal reasoning.The cited approaches improve different stages of the process, including interpreting user requests and grounding generation or reasoning with retrieved visual or factual information.
- Persistent representations: Existing methods demonstrate the value of agentic mediation but usually tailor intermediate representations to particular interventions rather than preserving semantic commitments throughout the generation lifecycle.Resolved information, verification outcomes, and repair decisions may therefore fail to remain reliably tied to the same underlying commitments.
- Skill abstractions: Skills are reusable procedural knowledge that extends language agents beyond one-off tool use and can be authored, distilled from demonstrations or trajectories, or selected from repositories for a task.Recent multimodal systems also use skill abstractions to support complex visual reasoning and generation workflows.
3 Method
SCOPE preserves semantic commitments through an evolving structured specification that coordinates generation, verification, and conditional skill invocation. Gen-Arena evaluates complex visual intent using entity- and constraint-level specifications with a strict entity-first pass rule.
- SCOPE design: Its fixed core pipeline is Decomposer →Synthesizer →Generator →Verifier, while retrieval, reasoning, and repair are conditionally orchestrated around commitment-level issues.The Decomposer creates the specification, the Synthesizer forms the generation prompt, the Generator produces images, and the Verifier checks entities and constraints item by item.
- SCOPE design: SCOPE represents visual intent as an evolving specification z = (E, C, U), containing entities, verifiable commitments, and unresolved information.The specification serves as a shared interface across generation, verification, and targeted skill invocation.
- Conditional orchestration: Retrieval addresses missing external evidence, reasoning resolves implicit or underspecified commitments, and repair corrects verified visual failures without additional grounding or reasoning.Each skill updates the same evolving specification and remains anchored to a concrete unknown or violated commitment.
- Verification and repair: Verification evaluates the explicit item set I = E ∪ C and routes failures either to continued semantic resolution or targeted visual repair.Repair selects prompt rewriting, image editing, or regeneration based on whether the failure concerns prompt expression, localized defects, or broad entangled errors.
- Gen-Arena: Gen-Arena pairs natural-language prompts with specifications identifying required entities and their constraints, enabling commitment-level evaluation across six manually annotated categories.The benchmark includes cartoon, game, sports, entertainment, competition, and ceremony instances, with references collected when text cannot reliably specify identity or appearance.
- Gen-Arena: EGIP applies an entity-first strict pass rule: an instance passes only when all required entities are correctly realized and all associated constraints are satisfied.Any missing or incorrectly depicted required entity causes failure before constraint checking proceeds.
4 Experiments
Experiments evaluate SCOPE on commitment-level realization in Gen-Arena and transfer to WISE-V and MindBench. SCOPE substantially improves strict intent fulfillment and external benchmark performance, with ablations attributing gains to coordinated retrieval, reasoning, and repair around persistent specifications.
- Evaluation setup: SCOPE is evaluated on Gen-Arena using EGIP across six categories and on WISE-V and MindBench for external transfer.The study compares SCOPE with Nano Banana Pro and multiple open-source image-generation models.
- Gen-Arena results: SCOPE surpasses Nano Banana Pro by 39 percentage points in overall Gen-Arena EGIP, with consistent gains across all six categories.The strongest improvements occur in Sports and Ceremony, which require identity grounding, event-specific relations, and precise scene composition.
- External benchmark transfer: 0.907 overall WiScore is achieved by SCOPE on WISE-V, where it ranks first in five of six categories and improves over Nano Banana Pro by 3.5% overall.These results support maintaining retrieved evidence, inferred requirements, and verification feedback within one structured specification.
- External benchmark transfer: 0.61 overall accuracy is achieved by SCOPE on MindBench, alongside 0.63 in Reasoning and a 48.8% improvement over Nano Banana Pro in overall accuracy.The result further corroborates SCOPE’s transfer to complex generation benchmarks.
- Ablation analysis: 0.40 EGIP for Direct (best-of-3) exceeds 0.21 for Direct (single), while Self-refine w/o spec reaches only 0.39 under the same three-generation budget.Removing retrieval and reasoning reduces SCOPE to 0.22 EGIP, indicating that structured decomposition alone is insufficient without semantic resolution.
- Diagnostic analysis: 0.82 item-level Entity Pass for Nano Banana Pro can still yield low EGIP because one missing or incorrectly grounded entity blocks dependent attributes, relations, or layouts.EGIP requires every required entity in an instance to be correct before its constraints receive credit.
5 Conclusion
SCOPE addresses the Conceptual Rift by persistently representing semantic commitments in an evolving specification and conditionally orchestrating skills around unresolved or violated commitments. The paper also introduces Gen-Arena and EGIP for commitment-level evaluation.
- Framework: SCOPE represents complex visual-intent commitments in an evolving structured specification to address the Conceptual Rift.The framework tracks commitments as persistent operational units across the generation lifecycle.
- Framework: SCOPE orchestrates retrieval, reasoning, and repair skills around commitments that remain unresolved or violated.Skill invocation is conditional on commitment status within the evolving specification.
- Evaluation: Gen-Arena is a human-annotated benchmark for commitment-level evaluation, paired with EGIP as a strict entity-first pass metric.Together, they evaluate whether generated images satisfy entity- and commitment-level intent.
Limitations
SCOPE improves commitment-level intent realization but incurs higher inference-time cost and latency than one-shot generation. It also depends on item-level verification to route subsequent repair actions.
- Computational overhead: SCOPE requires multiple calls across up to three iterations per instance, increasing inference-time cost and latency relative to one-shot generation.Calls may involve the MLLM, image generator, verifier, and optionally retrieval services.
- Computational overhead: Adaptive early stopping or more selective skill invocation could reduce SCOPE’s inference overhead.
- Verification dependence: SCOPE depends on item-level verification to route subsequent repair actions, making verification a limitation of the framework.