GenAgent: Scaling Text-to-Image Generation via Agentic Multimodal Reasoning
GenAgent employs agentic multimodal reasoning with tool invocation, achieving +23.6% performance on GenEval++.
Key Findings
Methodology
GenAgent utilizes a two-stage training process: first, supervised fine-tuning (SFT) on high-quality tool invocation and reflection data to bootstrap behaviors; second, end-to-end agentic reinforcement learning (RL) with hybrid rewards—pointwise for final image quality and pairwise for reflection consistency—plus trajectory resampling to enhance multi-turn exploration. The core architecture is a multimodal reasoning chain enabling autonomous planning, tool invocation, output evaluation, and reflection across multiple interaction rounds, forming a dynamic decision-making loop.
Key Results
- Performance on GenEval++ improved by 23.6%, from 0.325 to 0.507; on WISE, a 14% increase to 0.69. Using Qwen-Image as the external tool, performance nears GPT-4o (0.72 vs. 0.80), demonstrating strong cross-tool generalization. Multi-turn interactions progressively refine outputs, with the model automatically adapting reasoning strategies to different tasks.
- Compared to end-to-end and static modular systems, GenAgent surpasses baselines in complex reasoning and creative generation, validating the effectiveness of multi-round autonomous reasoning. Ablation studies confirm the importance of training stages and reward mechanisms in performance gains.
- The hybrid reward system effectively guides the model to improve both final output and intermediate reflections, while trajectory resampling broadens exploration, leading to robust multi-turn decision-making. Cross-tool adaptability and task-specific reasoning emerge naturally, indicating high scalability and flexibility.
Significance
This work advances the field by decoupling understanding and generation via an agentic framework, enabling autonomous multi-turn reasoning with external tools. It addresses longstanding challenges of high training costs and limited flexibility in unified models, offering a scalable, interpretable, and adaptable approach. The ability to generalize across tools and tasks, combined with iterative refinement, paves the way for more autonomous, intelligent multimodal systems with broad applications in content creation, virtual environments, and AI-assisted design. Its innovative architecture and training paradigm set new benchmarks for scalable, flexible AI capable of complex reasoning and creative synthesis.
Technical Contribution
The paper introduces a novel agentic multimodal architecture that decouples understanding from generation, leveraging external image generation tools as callable modules within a multi-turn reasoning loop. The two-stage training combines supervised fine-tuning with high-quality synthetic trajectories and reinforcement learning with hybrid rewards, employing trajectory resampling to promote exploration. This design enables the model to generalize across tools with varying capabilities, adapt reasoning strategies automatically to different tasks, and scale performance during inference without retraining. The hybrid reward mechanism and multi-round trajectory resampling are key innovations that enhance exploration and learning efficiency, setting a new paradigm for scalable, flexible multimodal AI systems.
Novelty
This is the first work to embed agentic, multi-turn reasoning with external tool invocation into text-to-image generation, moving beyond static pipelines and monolithic models. Unlike prior approaches that rely solely on end-to-end training or fixed workflows, GenAgent dynamically plans, invokes tools, and refines outputs through iterative reflection, demonstrating superior adaptability and scalability. Its hybrid reward strategy and trajectory resampling further distinguish it from existing methods, enabling effective multi-round exploration and task-specific reasoning, thus representing a significant leap in multimodal AI design.
Limitations
- Despite strong performance, the model's multi-round reasoning can still be affected by inaccuracies in tool invocation and reflection, especially in highly complex or ambiguous scenarios, limiting ultimate robustness.
- Training relies heavily on synthetic, curated datasets which are costly to produce, and the generalization to real-world, noisy data remains to be fully validated.
- Computational costs are non-trivial due to multi-round inference and reinforcement learning, posing challenges for deployment at scale in resource-constrained environments.
Future Work
Future research will focus on improving the robustness of tool invocation and reflection, reducing training data costs through better data augmentation, and extending the framework to incorporate more external tools (e.g., text, audio). Additionally, exploring multi-task learning and continual learning paradigms could further enhance task adaptability and efficiency, pushing towards more autonomous, general-purpose multimodal reasoning systems.
AI Executive Summary
In the rapidly evolving field of multimodal AI, the challenge of integrating understanding and generation remains a central hurdle. Traditional models, whether end-to-end or modular, face trade-offs between performance, flexibility, and scalability. End-to-end systems like Florance or Gato excel in certain tasks but demand enormous training resources and often lack interpretability. Modular systems, such as PromptEnhancer or ReflectionFlow, offer flexibility but are limited by static workflows that cannot adapt dynamically to complex scenarios.
This paper introduces GenAgent, a pioneering framework that unifies visual understanding and image generation through an agentic multimodal reasoning architecture. Unlike conventional models, GenAgent decouples understanding from generation, treating image generators as invokable external tools within a multi-turn reasoning loop. The core idea is to enable the model to autonomously plan, invoke tools, evaluate outputs, and reflect iteratively, forming a chain-of-thought process that enhances output quality progressively.
The training strategy combines supervised fine-tuning on synthetic, high-quality trajectories with reinforcement learning guided by hybrid rewards—pointwise for final image quality and pairwise for reflection consistency. Trajectory resampling during training encourages diverse exploration across multiple interaction rounds. Experimental results on benchmarks like GenEval++ and WISE demonstrate performance boosts of 23.6% and 14%, respectively, with performance approaching that of GPT-4o when using Qwen-Image as the external generator.
Beyond raw performance, GenAgent exhibits three key properties: cross-tool generalization, test-time scaling with multi-round refinement, and task-adaptive reasoning. These features collectively enable the model to adapt seamlessly to different tools and tasks, significantly advancing the scalability and flexibility of multimodal AI systems. The framework’s interpretability and modularity open new avenues for future research, including expanding external tool integration and optimizing multi-turn reasoning.
Overall, this work marks a significant step toward autonomous, scalable, and interpretable multimodal AI, with broad implications for content creation, virtual reality, and intelligent design. The authors’ approach demonstrates that decoupling understanding and generation, combined with multi-turn reasoning, can unlock new levels of performance and adaptability in AI systems, setting a new paradigm for future multimodal research.
Deep Dive
Abstract
We introduce GenAgent, unifying visual understanding and generation through an agentic multimodal model. Unlike unified models that face expensive training costs and understanding-generation trade-offs, GenAgent decouples these capabilities through an agentic framework: understanding is handled by the multimodal model itself, while generation is achieved by treating image generation models as invokable tools. Crucially, unlike existing modular systems constrained by static pipelines, this design enables autonomous multi-turn interactions where the agent generates multimodal chains-of-thought encompassing reasoning, tool invocation, judgment, and reflection to iteratively refine outputs. We employ a two-stage training strategy: first, cold-start with supervised fine-tuning on high-quality tool invocation and reflection data to bootstrap agent behaviors; second, end-to-end agentic reinforcement learning combining pointwise rewards (final image quality) and pairwise rewards (reflection accuracy), with trajectory resampling for enhanced multi-turn exploration. GenAgent significantly boosts base generator(FLUX.1-dev) performance on GenEval++ (+23.6\%) and WISE (+14\%). Beyond performance gains, our framework demonstrates three key properties: 1) cross-tool generalization to generators with varying capabilities, 2) test-time scaling with consistent improvements across interaction rounds, and 3) task-adaptive reasoning that automatically adjusts to different tasks. Our code will be available at \href{https://github.com/deep-kaixun/GenAgent}{this url}.