Unified Thinker: A General Reasoning Modular Core for Image Generation

TL;DR

The paper addresses the reasoning–execution gap in open-source image generation by introducing Unified Thinker, a decoupled think-then-execute core that grounds planning in the capabilities of a diffusion-based Generator. It introduces HieraReason-40K, a ~40K dataset of structured reasoning traces and executable prompts, and a two-stage training regime consisting of joint supervised fine-tuning and execution-guided dual-phase reinforcement learning to align planning with pixel-level outcomes. Empirical results show substantial gains in reasoning-intensive generation and editing across multiple benchmarks and demonstrate cross-generator transferability, supporting modular, generator-agnostic reasoning for T2I and editing tasks. The work advances practical, instruction-following image generation by separating reasoning from rendering, enabling reusable, executable planning across diverse backbones and workflows.

Abstract

Despite impressive progress in high-fidelity image synthesis, generative models still struggle with logic-intensive instruction following, exposing a persistent reasoning--execution gap. Meanwhile, closed-source systems (e.g., Nano Banana) have demonstrated strong reasoning-driven image generation, highlighting a substantial gap to current open-source models. We argue that closing this gap requires not merely better visual generators, but executable reasoning: decomposing high-level intents into grounded, verifiable plans that directly steer the generative process. To this end, we propose Unified Thinker, a task-agnostic reasoning architecture for general image generation, designed as a unified planning core that can plug into diverse generators and workflows. Unified Thinker decouples a dedicated Thinker from the image Generator, enabling modular upgrades of reasoning without retraining the entire generative model. We further introduce a two-stage training paradigm: we first build a structured planning interface for the Thinker, then apply reinforcement learning to ground its policy in pixel-level feedback, encouraging plans that optimize visual correctness over textual plausibility. Extensive experiments on text-to-image generation and image editing show that Unified Thinker substantially improves image reasoning and generation quality.

Figures (9)

• Figure 1: Challenges in reasoning-aware image generation. Existing models, exemplified by Qwen-Image-Edit, exhibit two failure modes: (1) inaccurate reasoning (without Thinker), leading to logically incorrect edits; and (2) imprecise rendering (with Thinker), where correct reasoning does not translate into faithful visual outputs. Our Unified Thinker aims to address both issues.
• Figure 2: Visual demonstrations of Unified Thinker on unified image generative tasks, including image editing and text-to-image generation, along with reasoning.
• Figure 3: Data construction pipeline for HieraReason-40K. We combine seed knowledge and user requests to generate structured reasoning traces and executable enhanced prompts.
• Figure 4: Our proposed two-stage framework for reasoning-aware image generation. Stage 1 initializes the Thinker Model and Generator Model. Given an Image & Prompt (x), the Thinker generates a Reasoning Thought (y), which then guides the Generator to produce a Refined Image (z). Stage 2 further refines the Thinker and Generator Models to enhance their capability in integrating complex reasoning (y) into high-fidelity visual outputs (z), applicable to both novel image generation and existing image editing tasks.
• Figure 5: Mean reward score over training.