SCOPE: Structured Decomposition and Conditional Skill Orchestration for Complex Image Generation

TL;DR

SCOPE is a specification-guided orchestration framework for complex text-to-image generation: every prompt is parsed into a structured spec of entities, atomic constraints, and unknowns, and that spec is the persistent state passed between a Decomposer, retrieval/reasoning skills, the generator, a Verifier, and targeted repair skills. The framing concept is the Conceptual Rift — that multi-step generation pipelines lose track of which semantic commitment a given retrieval/verification/repair operation is resolving, so failures bleed across steps. The paper also introduces Gen-Arena, a 300-instance human-annotated benchmark with entity- and constraint-level ground truth, plus EGIP (Entity-Gated Intent Pass Rate), a strict pass criterion that requires all required entities and all gated constraints to be satisfied. SCOPE reaches 0.60 EGIP on Gen-Arena versus 0.21 for the strongest baseline (Nano Banana Pro) and 0.01–0.02 for Qwen-Image, Z-Image-Turbo, and FLUX.1-dev.

Project page: https://nopnor.github.io/SCOPE/. Code (MIT): https://github.com/nopnor/SCOPE.

Key claims

Multi-step image-generation pipelines (retrieval, generation, verification, repair) suffer from a Conceptual Rift: each operation locally resolves a fact but the operational unit — which commitment was just satisfied — is not preserved across the lifecycle, so commitments fail to remain identifiable throughout [Abstract, §1].
The fix is a single persistent structured specification — explicit lists of required entities, atomic constraints (attribute / relation / layout), and unresolved unknowns — that is the shared interface for all skills in the pipeline; retrieval, reasoning, and repair are invoked conditionally on what the current spec exposes as unresolved or violated, rather than unconditionally on every step [§Method, project page].
The Verifier maps item-level failures back to spec entries so that the repair skills can target the specific violated commitment rather than regenerate from scratch [§Method, project page].
Gen-Arena: 300 human-annotated instances (cartoon, game, sports, entertainment, competition, ceremony), 1,954 required entities, 2,533 atomic constraints, 310 reference images [project page §Gen-Arena].
EGIP (Entity-Gated Intent Pass Rate) is the strict metric: an instance passes only if every required entity and every gated constraint is satisfied — designed to expose the gap between entity recall (which baselines hit easily) and full-spec fulfillment [project page §EGIP].
SCOPE reaches 0.60 EGIP overall on Gen-Arena vs. 0.21 for Nano Banana Pro, 0.07 for Nano Banana, 0.02 for Qwen-Image, 0.01 for Z-Image-Turbo, 0.01 for FLUX.1-dev [project page Results table].
Entity recall is already high for baselines (Nano Banana Pro 0.82, Qwen-Image 0.83, Z-Image-Turbo 0.84); the bottleneck baselines miss is gated constraint satisfaction, where SCOPE goes 0.83 vs Nano Banana Pro 0.59 / Qwen-Image 0.49 / Z-Image-Turbo 0.48 [project page Entity vs. Gated Constraint table].
Ablation: removing retrieval+reasoning drops EGIP 0.60 → 0.22 (collapsing to roughly the Direct-single 0.21 baseline); removing repair drops EGIP 0.60 → 0.42; best-of-3 direct generation reaches only 0.40 — so the repair loop accounts for ~0.18 EGIP and the retrieval/reasoning stage accounts for ~0.38, with neither replaceable by additional sampling [project page Ablation table].
SCOPE transfers to external benchmarks: 0.907 on WISE-V (vs. best baseline 0.876) and 0.61 on MindBench Overall (vs. 0.41 best baseline, with MindBench Knowledge 0.59 vs. 0.40 and MindBench Reasoning 0.63 vs. 0.44) [project page External benchmarks table].

Method

SCOPE structures complex image generation as a closed loop over an evolving specification rather than a single forward pass. The Decomposer parses the prompt into entities, atomic constraints (attribute, relation, layout), and unknowns. When the spec exposes missing evidence or unresolved requirements, SCOPE invokes retrieval (looking up reference imagery / facts about a named entity) and reasoning skills to fill them before generation. The generator then renders an image; the Verifier checks the image against each spec item and maps failures back to specific commitments. Repair skills target only those violated commitments — they do not regenerate the full image blindly. The loop iterates until the spec is satisfied or a budget is hit.

The structured spec is the load-bearing piece: it is the persistent state that lets each later operation know which semantic commitment it is addressing, defeating the Conceptual Rift the paper names.

Results

Gen-Arena EGIP (overall): SCOPE 0.60 vs. Nano Banana Pro 0.21, Nano Banana 0.07, Qwen-Image 0.02, Z-Image-Turbo 0.01, FLUX.1-dev 0.01 [project page Results].
Entity vs. Gated Constraint breakdown: SCOPE 0.92 entity / 0.83 gated constraint; Nano Banana Pro 0.82 / 0.59; Qwen-Image 0.83 / 0.49; Z-Image-Turbo 0.84 / 0.48 — confirming the gap is constraint satisfaction, not entity recall [project page].
Ablations: Direct single 0.21 EGIP / 0.59 gated; Direct best-of-3 0.40 / 0.71; SCOPE w/o retrieval+reasoning 0.22 / 0.60; SCOPE w/o repair 0.42 / 0.72; full SCOPE 0.60 / 0.83 [project page].
External benchmarks: WISE-V Overall 0.907 (vs. 0.876 best baseline); MindBench Knowledge 0.59 (vs. 0.40), Reasoning 0.63 (vs. 0.44), Overall 0.61 (vs. 0.41) [project page].

Why it’s interesting

SCOPE is the cleanest filed instance of the “VLM scores intermediate visuals in a chain of reasoning, and the score is used to drive a targeted repair action” pattern — UniG2U-Bench (UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?) named the scoring half of this loop and IGenBench (IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation) introduced the prompt-decomposed atomic-yes/no rubric, but neither closes the loop into repair. SCOPE does, and the ablation (0.18 EGIP attributable to repair, 0.38 to retrieval+reasoning) is the first filed quantification of how much each stage of the verify-and-repair loop contributes on a strict EGIP-style metric. Adjacent to Tool-Use Agents in that the persistent spec plays the same role a ReAct scratchpad plays for text agents, but with constraints as first-class objects rather than free-text observations — closer in spirit to MiroThinker’s structured trajectory (MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling) than to a free-form CoT. As a seventh axis of Inference-Time Scaling, SCOPE spends inference compute on a verification-repair loop driven by a structured spec — orthogonal to KV-cache compression, scaffold-side sub-LM dispatch, NFE×K refinement, and TTT weight updates.