[Paper] UniReason 1.0: A Unified Reasoning Framework for World Knowledge Aligned Image Generation and Editing

Source: arXiv - 2602.02437v1

Overview

UniReason 1.0 tackles a long‑standing gap in multimodal AI: the disconnect between text‑to‑image generation and image editing. By treating both as linked reasoning steps—first planning a scene with world knowledge, then refining it through self‑reflection—the authors deliver a single model that can both imagine and polish images in a human‑like “plan‑then‑fix” workflow.

Key Contributions

• Dual‑reasoning framework that unifies generation (knowledge‑driven planning) and editing (visual self‑correction) under a shared latent representation.
• Reasoning‑centric dataset (~300 k samples) spanning five knowledge domains (cultural commonsense, physics, geometry, everyday logic, and temporal relations) to teach the model how to plan coherent scenes.
• Agent‑generated self‑correction corpus that provides examples of visual errors and the corresponding edits, enabling the model to learn “self‑reflection.”
• State‑of‑the‑art results on reasoning‑heavy benchmarks (WISE, KrisBench, UniREditBench) while retaining strong performance on standard synthesis tasks.
• Open‑source implementation (code and data) that encourages further research on unified generative‑editing pipelines.

Methodology

1. Shared Representation Layer – Both generation and editing modules feed into a common transformer‑based latent space, allowing knowledge and visual cues to be exchanged freely.
2. World‑Knowledge‑Enhanced Planning – The model first parses the textual prompt, retrieves relevant facts from a curated knowledge base, and produces a high‑level “plan” (e.g., object layout, physical constraints). This plan guides the initial image synthesis.
3. Self‑Reflection Editing – After the first image is rendered, a lightweight visual critic (trained on the self‑correction corpus) detects inconsistencies (e.g., a floating object, wrong lighting) and proposes pixel‑level edits. The editing module iteratively refines the image until the visual critic signals convergence.
4. Training Regime – The system is trained end‑to‑end on the combined dataset: the planning branch learns from the reasoning‑centric samples, while the editing branch learns from the agent‑generated correction pairs. A multi‑task loss balances semantic fidelity, visual realism, and logical consistency.

Results & Findings

Interpretation: UniReason dramatically narrows the performance gap on tasks that require deep reasoning, while staying competitive on classic generation metrics. Qualitative examples show the model correctly placing objects according to physical laws (e.g., a cup resting on a table) and fixing subtle errors like mismatched shadows after the initial render.

Practical Implications

• Content Creation Pipelines – Designers can issue a single prompt (“A medieval market at dusk with realistic lighting”) and receive a plausibly composed image that the system automatically polishes, cutting down on manual touch‑ups.
• Interactive Editing Tools – Developers can embed UniReason into photo‑editing software to enable “smart fix” actions: the user flags a visual inconsistency, and the model suggests a context‑aware correction.
• Simulation & Training Data – Synthetic datasets for robotics or AR can be generated with built‑in physical consistency, reducing the need for costly manual validation.
• Explainable AI – Because the model produces an explicit planning graph before rendering, developers can inspect the reasoning chain (e.g., “object A must be on surface B”) to debug or enforce domain‑specific constraints.
• Cross‑Domain Consistency – Applications that span multiple modalities (e.g., generating an illustration for a technical manual) benefit from the unified knowledge base that aligns visual output with factual content.

Limitations & Future Work

• Knowledge Base Scope – The current reasoning corpus covers five domains; extending to specialized fields (medical, legal) will require additional curated data.
• Computation Overhead – The two‑stage plan‑then‑edit loop incurs higher latency than single‑pass generators, which may be a bottleneck for real‑time applications.
• Error Propagation – Mistakes in the planning stage can sometimes mislead the editing module, leading to sub‑optimal refinements. Future work aims to incorporate a feedback loop where the visual critic can request replanning.
• Evaluation Diversity – Benchmarks focus on static images; exploring video generation/editing with temporal reasoning is an open direction.

Bottom line: UniReason 1.0 demonstrates that unifying generation and editing through structured reasoning is not just a research curiosity—it’s a practical step toward more intelligent, self‑correcting visual AI that can be plugged into everyday developer toolchains.

Authors

• Dianyi Wang
• Chaofan Ma
• Feng Han
• Size Wu
• Wei Song
• Yibin Wang
• Zhixiong Zhang
• Tianhang Wang
• Siyuan Wang
• Zhongyu Wei
• Jiaqi Wang

Paper Information

• arXiv ID: 2602.02437v1
• Categories: cs.CV, cs.AI
• Published: February 2, 2026
• PDF: Download PDF