R³ Loop: Enabling Self-Reflection and Correction in AI Image Generation

The CUHK team proposes the Reason-Reflect-Rectify (R³) framework, breaking through the bottleneck of the single-generation paradigm in text-to-image (T2I) models; constructs the R³-Bench evaluation benchmark to reveal the capability gap of current models—"can identify problems but cannot correct them"; and presents the R³-Refiner two-stage optimization framework, which achieves a 12% increase in reflection judgment score and a 9% increase in correction score, while also having cross-model compatibility.

Current mainstream text-to-image (T2I) and unified multimodal models (UMMs) rely on a single-generation paradigm: after users input prompts, the model directly outputs images. This mode struggles to meet requirements in one go when handling complex prompts (such as specific spatial relationships, quantity constraints, or style combinations). When users find issues, they can only regenerate images without targeted improvements.

The R³ Loop consists of three stages:

1. Reason: Analyze the deep semantic needs of prompts and identify key constraints;
2. Reflect: Examine generated results and judge discrepancies from prompts;
3. Rectify: Generate specific executable correction instructions to guide the next round of generation. The three stages form a closed loop, allowing the model to approach user expectations through multi-round iterations.

The research team constructed the R³-Bench benchmark dataset (containing over 600 expert-annotated instances) to evaluate models based on reflection judgment score (ability to identify errors) and correction score (ability to generate executable instructions). The results show that current state-of-the-art models can identify errors but cannot generate actionable correction instructions, presenting a core bottleneck of "can find problems but cannot solve them".

R³-Refiner is a two-stage framework based on reinforcement learning:

• Stage 1: Group Relative Policy Optimization (GRPO), a reinforcement learning algorithm without a value network, training high-quality reflection and correction strategies;
• Stage 2: Hierarchical Reward Mechanism (HRM), a layered reward structure (semantic consistency, executability, effect verification rewards) to ensure the effectiveness of correction instructions.

R³-Refiner achieves on R³-Bench: a 12.0% increase in reflection judgment score and a 9.0% increase in correction score; it has cross-model compatibility and can be integrated into various multimodal large language models (MLLMs) and T2I models (such as the Stable Diffusion series). Its performance in following complex prompts on benchmarks like GenEval++ and T2I-CompBench is better than the baseline.

The R³ framework marks a paradigm shift in text-to-image generation from "single-generation" to "iterative optimization":

• Application scenarios: designers' multi-round refinement of concept maps, complex scene generation, model capability diagnosis;
• Open source: the code has been open-sourced (https://github.com/xiaomoguhz/R3-Bench);
• Future: expand to video/3D generation, explore human-machine collaborative interactive generation modes.