Generation Navigator: A State-Aware Agentic Framework for Image Generation
Generation Navigator optimizes image generation with PRE-GRPO, achieving a WISE score of 0.90.
Key Findings
Methodology
This paper introduces the Generation Navigator framework, redefining image generation as a state-conditioned action-making problem. By incorporating the PRE-GRPO algorithm, the framework dynamically adjusts the generation trajectory during multi-turn interactions. PRE-GRPO optimizes the process by rewarding high-quality image discovery, avoiding quality degradation, and minimizing unnecessary turns.
Key Results
- On T2I-ReasonBench, Generation Navigator achieved a reasoning accuracy of 79.06%, significantly outperforming existing methods.
- On the WISE benchmark, the method achieved an overall score of 0.90, surpassing many proprietary models.
- Ablation studies showed that PRE-GRPO exhibits more significant performance improvements compared to traditional GRPO algorithms.
Significance
Generation Navigator addresses the challenge of faithfully realizing user intent in image generation by introducing state-aware action decision-making. This framework not only enhances the quality of generated images but also reduces user interaction, holding significant academic and industrial value.
Technical Contribution
Technically, the paper achieves optimization of multi-turn generation trajectories through the PRE-GRPO algorithm, overcoming the credit assignment challenge in traditional reinforcement learning. This method introduces new theoretical guarantees and engineering possibilities in the field of image generation.
Novelty
This is the first approach to redefine image generation as a state-conditioned action-making problem and optimize generation trajectories with the PRE-GRPO algorithm, contrasting sharply with existing fixed workflow methods.
Limitations
- The method may still require multiple interactions to achieve satisfactory results when handling extremely complex user intents.
- In certain scenarios, the quality of generation may be limited by the accuracy of the initial prompt.
Future Work
Future research could explore further reducing interaction counts and improving generation quality. Additionally, extending the framework to support more types of multimodal inputs is a promising direction.
AI Executive Summary
Despite rapid advances in text-to-image generation, accurately realizing user intent remains challenging. Existing methods often rely on simple prompt rewriting or rule-based closed-loop agents, failing to adapt dynamically to the evolving generation process.
This paper proposes the Generation Navigator framework, redefining image generation as a state-conditioned action-making problem. By incorporating the PRE-GRPO algorithm, the framework dynamically adjusts the generation trajectory during multi-turn interactions, significantly enhancing the quality and efficiency of generated images.
Experimental results show that Generation Navigator performs exceptionally well across multiple benchmarks, particularly achieving significant performance improvements on T2I-ReasonBench and WISE. This framework holds significant academic importance and opens new possibilities for industrial applications.
Deep Analysis
Background
In recent years, text-to-image generation has made significant progress, particularly in diffusion models and autoregressive multimodal generation. However, single-pass generation often struggles to meet complex user intents, especially when involving spatial layouts, commonsense reasoning, and fine-grained styles.
Core Problem
Existing methods often rely on fixed workflows or prompt rewriting, unable to dynamically adapt to the changing generation process. This results in suboptimal outcomes, requiring users to make multiple attempts to achieve satisfactory results.
Innovation
Generation Navigator redefines image generation as a state-conditioned action-making problem, allowing dynamic adjustment of the generation process. The PRE-GRPO algorithm optimizes the generation trajectory by rewarding high-quality image discovery, avoiding quality degradation, and minimizing unnecessary turns.
Methodology
- �� The Generation Navigator framework consists of a navigator, generator, and reviewer. • The navigator is the only learned component responsible for outputting structured actions. • The PRE-GRPO algorithm optimizes the generation trajectory by rewarding high-quality image discovery, avoiding quality degradation, and minimizing unnecessary turns.
Experiments
Experiments were conducted on benchmarks like T2I-ReasonBench and WISE, using Qwen3-VL-8B-Instruct as the navigator, Doubao-Seed1.5 as the reviewer, and FLUX.2-Klein-9B as the generator. Results show that Generation Navigator performs exceptionally well across multiple benchmarks.
Results
On T2I-ReasonBench, Generation Navigator achieved a reasoning accuracy of 79.06%, significantly outperforming existing methods. On the WISE benchmark, the method achieved an overall score of 0.90, surpassing many proprietary models.
Applications
This framework can be applied in scenarios requiring high-quality image generation, such as advertising creativity and film production. Its state-aware action decision-making mechanism significantly improves generation efficiency and reduces user interaction.
Limitations & Outlook
Despite its excellent performance across benchmarks, the method may still require multiple interactions to handle extremely complex user intents. Additionally, the quality of generation may be limited by the accuracy of the initial prompt.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. You have a recipe (user prompt), but each dish you make (generated image) doesn't taste quite right. So you start adjusting ingredients and steps (state-conditioned action decisions) until you create a satisfying dish. Generation Navigator is like a smart cooking assistant that adjusts the cooking steps based on each attempt's results, helping you make the perfect dish.
ELI14 Explained like you're 14
Imagine you're playing a game where you have to draw a picture each time. You have a theme, but each drawing doesn't quite match. So you start adjusting your drawing method, like changing colors or starting over. Generation Navigator is like a smart game assistant that gives you suggestions based on each drawing's result, helping you create the perfect artwork. Isn't that cool?
Glossary
Generation Navigator
A multi-turn text-to-image generation framework that dynamically adjusts the generation trajectory.
Used to optimize the image generation process.
PRE-GRPO
A trajectory-level reinforcement learning objective that rewards high-quality image discovery and minimizes unnecessary turns.
Used to optimize generation trajectories.
T2I-ReasonBench
A benchmark for evaluating text-to-image generation models' reasoning capabilities.
Used to validate the model's reasoning accuracy.
WISE
A knowledge-intensive generation benchmark evaluating models' generation capabilities across multiple domains.
Used to validate the model's knowledge generation capabilities.
Qwen3-VL-8B-Instruct
A multimodal language model used as the navigator.
Serves as a core component of Generation Navigator.
Open Questions Unanswered questions from this research
- 1 How to reduce interaction counts under extremely complex user intents remains an open question.
- 2 How to further improve generation quality, especially when initial prompts are inaccurate.
Applications
Immediate Applications
Advertising Creativity
Advertising companies can use this framework to generate high-quality ad images, reducing trial and error in the creative process.
Long-term Vision
Film Production
This framework can be used in scene design for film production, significantly improving production efficiency and image quality.
Abstract
Despite rapid advances in text-to-image generation, faithfully realizing user intent remains challenging, often requiring manual multi-turn trial and error. To automate this process, existing systems rely on either simple prompt rewriting or closed-loop agents driven by hand-crafted rules, rather than learning to adapt actions to the evolving generation process. In this paper, we reformulate image generation as a state-conditioned action-making problem and propose Generation Navigator, a multi-turn T2I agent that learns to dynamically steer the generation trajectory and output the next action. However, training this agent via reinforcement learning introduces a critical credit assignment challenge: naively rewarding a trajectory based solely on a single state assigns equal credit to all actions in the rollout, ignores the quality dynamics across turns, and fails to distinguish actions that improve the trajectory from those that degrade it or waste turns without progress. We resolve this with PRE-GRPO (Peak-Retention-Efficiency Group Relative Policy Optimization), a trajectory-level reinforcement learning objective that explicitly rewards discovering a high-quality image (Peak), avoiding subsequent quality degradation across turns (Retention), and minimizing unnecessary turns (Efficiency). Experiments show substantial improvements across benchmarks, reaching a WISE score of 0.90 and 79.06% reasoning accuracy on T2I-ReasonBench.