Mind-Brush: Integrating Agentic Cognitive Search and Reasoning into Image Generation

TL;DR

The paper tackles the gap between static text-to-image decoders and the need for real-time knowledge and complex reasoning in visual synthesis. It introduces Mind-Brush, a unified agentic framework that couples intent analysis, multimodal retrieval, and explicit chain-of-thought reasoning to enable a Think-Research-Create workflow for image generation. To evaluate cognitive-generation capabilities, it presents Mind-Bench, a 500-sample benchmark across knowledge-driven and reasoning-driven tasks with a Checklist-based Strict Accuracy metric. Experimental results show that Mind-Brush significantly improves generation accuracy and grounding across Mind-Bench, WISE, and RISEBench, demonstrating robust open-world knowledge integration and reasoning for high-fidelity visual synthesis.

Abstract

While text-to-image generation has achieved unprecedented fidelity, the vast majority of existing models function fundamentally as static text-to-pixel decoders. Consequently, they often fail to grasp implicit user intentions. Although emerging unified understanding-generation models have improved intent comprehension, they still struggle to accomplish tasks involving complex knowledge reasoning within a single model. Moreover, constrained by static internal priors, these models remain unable to adapt to the evolving dynamics of the real world. To bridge these gaps, we introduce Mind-Brush, a unified agentic framework that transforms generation into a dynamic, knowledge-driven workflow. Simulating a human-like 'think-research-create' paradigm, Mind-Brush actively retrieves multimodal evidence to ground out-of-distribution concepts and employs reasoning tools to resolve implicit visual constraints. To rigorously evaluate these capabilities, we propose Mind-Bench, a comprehensive benchmark comprising 500 distinct samples spanning real-time news, emerging concepts, and domains such as mathematical and Geo-Reasoning. Extensive experiments demonstrate that Mind-Brush significantly enhances the capabilities of unified models, realizing a zero-to-one capability leap for the Qwen-Image baseline on Mind-Bench, while achieving superior results on established benchmarks like WISE and RISE.

Figures (24)

• Figure 1: We introduce Mind-Brush, an agentic framework that synergizes active search with explicit reasoning for image generation. By decomposing user intent, retrieving multimodal evidence, and inferring latent requirements, our agent effectively bridges the cognitive gaps and interpretative biases prevalent in existing models. Furthermore, we propose Mind-Bench, a comprehensive benchmark designed to evaluate model performance on up-to-date long-tail concepts and multimodal reasoning tasks, thereby probing the boundaries of unified understanding and generation capabilities.
• Figure 2: The overall framework of Mind-Brush. The user input first undergoes intent decomposition to identify potential knowledge deficits and formulate a question list. Based on specific requirements, the system dynamically executes specialized tools—such as utilizing active search or logical reasoning—to effectively bridge cognitive gaps. Finally, the consolidated evidence is organized into a final instruction via a concept review process to guide precise image generation, ensuring alignment with the user's authentic intent.
• Figure 3: Overview of Checklist-based Strict Accuracy (CSA) evaluation pipeline in Mind-Bench.
• Figure 4: Qualitative Comparison of different models on Mind-Bench, including knowledge-driven (upper part) and reasoning-driven (lower part) tasks. The green border indicates that the generated result matches the facts, while the red border indicates the opposite.
• Figure 5: A generation process of Mind-Brush in Special Events task of Mind-Bench.