UniReasoner: Bridging the Understanding-Generation Gap in Visual Generation Using the Reasoning Capabilities of Large Language Models

This paper proposes a formal definition of the understanding-generation gap and the UniReasoner framework. By using LLMs to generate visual drafts, perform self-critical evaluations, and output correction signals to guide diffusion models, it significantly improves compositional alignment and semantic fidelity while maintaining image quality.

Although text-to-image generation technology has made significant progress, there is a core paradox of "being able to understand but not draw correctly", such as missing attributes and incorrect relationships under complex prompts. The study formalizes this phenomenon as the "understanding-generation gap", whose causes include differences in conditional distributions between understanding and generation, discrete vs. continuous space mapping, and architectural flaws in one-way vs. two-way information flow.

The core of the UniReasoner framework is to use LLMs to convert understanding capabilities into generation guidance, consisting of three stages: 1. Visual draft generation (abstract representation of discrete visual tokens); 2. Self-critical evaluation (checking the consistency between the draft and the prompt, outputting correction signals); 3. Conditional diffusion generation (integrating three inputs: original prompt, visual draft, and text evaluation). Technical implementations include multi-scale visual tokenization, structured self-critical prompt engineering, and hierarchical diffusion condition fusion strategy.

Experiments show that UniReasoner significantly improves compositional alignment (spatial relationship accuracy from 62%→81%, attribute binding from 58%→76%) and semantic fidelity (prompt-image alignment improved by 23%, human preference rate of 65%), while maintaining image quality (equivalent FID scores, no difference in aesthetic quality). Ablation experiments prove that the combination of visual drafts and text evaluation produces a synergistic effect.

The reasons for UniReasoner's effectiveness include: 1. Explicit reasoning steps replace implicit learning, making the process visible and debuggable; 2. Reusing LLM verification capabilities to guide generation, forming a closed loop of "understanding→verification→guidance→generation"; 3. Hierarchical condition strategy applies signals at different abstract levels to achieve fine-grained control.

Current limitations: Increased computational overhead, risk of error accumulation, performance in complex scenarios needs improvement, domain generalization needs verification. Future directions: Iterative process optimization, expansion of interactive generation, multimodal applications, efficiency optimization, and adaptation to specific domains.

UniReasoner represents a new paradigm of "reasoning-driven generative AI", where generation includes explicit reasoning steps, understanding capabilities guide generation, and intermediate representations are interpretable and controllable. This paradigm can be extended to text, code, music generation, and other fields.

UniReasoner provides a practical path to bridge the understanding-generation gap, proving that LLM understanding capabilities can be converted into generation guidance without sacrificing quality. Core principle: Explicitly building a bridge between understanding and generation is key to constructing reliable and controllable AI systems. More AI systems that "think before creating" will emerge in the future.