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

Source: arXiv - 2605.08043v1

Overview

The paper introduces SCOPE, a new framework that lets text‑to‑image models keep track of every piece of a user’s visual intent—objects, attributes, spatial constraints, and more—throughout the whole generation process. By treating these intent pieces as semantic commitments and orchestrating specialized “skills” (retrieval, reasoning, repair) whenever a commitment is at risk, SCOPE dramatically improves the fidelity of complex image synthesis.

Key Contributions

• Commitment‑centric formulation – Defines semantic commitments and the “Conceptual Rift” problem where intent fragments get lost during generation.
• SCOPE architecture – A specification‑guided orchestration loop that maintains a structured, evolving spec and conditionally triggers retrieval, reasoning, and repair modules.
• Gen‑Arena benchmark – A human‑annotated dataset with fine‑grained entity and constraint specifications, plus the Entity‑Gated Intent Pass (EGIP) metric for strict entity‑first evaluation.
• State‑of‑the‑art results – SCOPE achieves 0.60 EGIP on Gen‑Arena, outpacing all baselines, and shows strong performance on existing suites (WISE‑V: 0.907, MindBench: 0.61).
• Open‑source components – The authors release the orchestration code and the Gen‑Arena benchmark, enabling reproducibility and further research.

Methodology

1. Structured Specification – The input prompt is parsed into a tree‑like spec containing entities (e.g., “red sports car”), attributes, and relational constraints (e.g., “behind a palm tree”).
2. Commitment Tracker – Each node in the spec becomes a commitment that is persisted across generation steps. The tracker flags commitments that are unresolved (not yet visualized) or violated (detected mismatch).
3. Conditional Skill Orchestration
   • Retrieval Skill – Pulls reference images or patches from a large visual database to provide concrete visual priors for a commitment.
   • Reasoning Skill – Uses a language‑vision model to infer missing details (e.g., “what does a vintage streetlamp look like?”) and to resolve ambiguous constraints.
   • Repair Skill – After an initial diffusion pass, a lightweight inpainting or refinement network edits the canvas to satisfy any violated commitments.
4. Iterative Loop – The system alternates between diffusion generation and skill invocation until all commitments are marked resolved or a maximum iteration budget is reached.
5. Evaluation – Gen‑Arena’s EGIP metric checks that every entity appears correctly before any other constraints are considered, ensuring a strict entity‑first success criterion.

Results & Findings

• Higher EGIP shows that SCOPE reliably renders every requested object, even when prompts contain 5‑10 entities with overlapping constraints.
• Qualitative analysis reveals fewer “conceptual rifts”: objects stay consistent across multi‑step generation, and spatial relationships (e.g., “to the left of”) are respected.
• Ablation studies confirm that each skill contributes: removing the repair module drops EGIP by ~0.12, while skipping retrieval reduces overall fidelity on rare objects.

Practical Implications

• Enterprise content creation – Marketing teams can feed highly detailed briefs (multiple products, brand colors, layout constraints) and obtain images that honor every element without manual post‑editing.
• Game asset pipelines – Designers can specify complex scene compositions (e.g., “a medieval market with a blacksmith beside a fountain”) and receive ready‑to‑use textures that respect spatial logic, cutting iteration time.
• E‑commerce – Automated generation of product‑in‑context shots (multiple items, specific lighting, background constraints) becomes feasible, reducing the need for costly photoshoots.
• Developer APIs – The orchestration loop can be exposed as a plug‑in for existing diffusion services (e.g., Stable Diffusion, DALL·E) to add “commitment tracking” as a service layer, enabling higher‑level control without retraining the base model.

Limitations & Future Work

• Scalability of the spec parser – Current rule‑based parsing struggles with highly ambiguous or colloquial prompts; a learned parser could improve robustness.
• Skill latency – Retrieval and reasoning steps add overhead (≈2–3 s per iteration), which may be prohibitive for real‑time applications. Optimizing these modules or caching common assets is an open direction.
• Generalization to unseen domains – The retrieval database is curated for common objects; rare or domain‑specific entities (e.g., medical equipment) still suffer from low fidelity. Expanding the database and incorporating domain‑adapted reasoning models are future targets.

Overall, SCOPE demonstrates that treating visual intent as a set of persistent commitments—and dynamically orchestrating specialized skills around them—can bridge the gap between human‑level specification and machine‑generated imagery.

Authors

• 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

Paper Information

• arXiv ID: 2605.08043v1
• Categories: cs.CV, cs.AI
• Published: May 8, 2026
• PDF: Download PDF