GenAgent: Scaling Text-to-Image Generation via Agentic Multimodal Reasoning

TL;DR

GenAgent presents an agentic multimodal framework that decouples visual understanding and generation by treating image generators as invokable tools, enabling autonomous multi-turn reasoning with multimodal chains-of-thought. It employs a two-stage training regime—cold-start supervised fine-tuning and agentic reinforcement learning with a hybrid reward and interaction-round resampling—to drive dynamic tool use and reflective refinement. Across GenEval++, WISE, and Imagine benchmarks, GenAgent with open-source tools achieves substantial gains and, with stronger generators like Qwen-Image, approaches GPT-4o performance while demonstrating cross-tool generalization and task-adaptive reasoning. This work offers a flexible, scalable approach to unify visual understanding and generation without prohibitive retraining costs, potentially transforming multimodal AI systems and their deployment in production settings.

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}.

Figures (7)

• Figure 1: The left shows the illustration of GenAgent's iterative interaction with image generation tools. The right presents the performance scaling of our GenAgent across interaction rounds (top) and generalization to different image generation tools (bottom).
• Figure 2: Overview of GenAgent. Upon receiving user input, GenAgent executes a multi-modal reasoning trace that encompasses iterative reasoning and image generation through dynamic interaction with image generation tools. During the rollout of agentic reinforcement learning, we compute pointwise rewards from the final image and pairwise rewards based on consecutive image pairs throughout trace.
• Figure 3: SFT data construction pipeline.
• Figure 4: Ablation study on image generation tools.
• Figure 5: Ablation on interaction rounds, with normalized scores.