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

Abstract

We introduce GenAgent, an agentic framework that unifies visual understanding and generation. GenAgent overcomes the limitations of tightly coupled unified models — which suffer from understanding–generation trade-offs and high training data costs — and the structural bottlenecks of existing complex multi-model pipelines through a flexible, decoupled architecture: understanding is handled by a single multimodal model, while generation is achieved by treating image generation models as invokable tools. Given an image prompt, GenAgent adaptively engages in reasoning, tool invocation, visual judgment, and reflection to iteratively refine outputs until the criteria are met. We employ a two-stage training strategy: first, we cold-start the agent with supervised fine-tuning on high-quality tool-invocation and reflection data to bootstrap multi-turn agent behaviors; second, we perform end-to-end agentic reinforcement learning, combining pointwise rewards (final image quality) and pairwise rewards (reflection accuracy), and utilizing a round-aware trajectory resampling strategy to balance improvements across distinct capabilities. Without modifying the underlying generators, GenAgent significantly boosts the base model (FLUX.1-dev) by +23.6% on GenEval++ and +14.0% on WISE. Beyond substantial performance gains, our framework demonstrates three key emergent 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.

+23.6%

GenEval++ Gain

+14.0%

WISE Gain

32K

SFT Trajectories

7B

Agent Backbone

3

Interaction Rounds

🧩

Decoupled Agentic Architecture

A single multimodal model orchestrates planning, judgment, and reflection, treating off-the-shelf image generators as callable tools — no interleaved data alignment required.

🎓

Two-Stage Training

Hint-guided SFT bootstraps multi-turn tool invocation and reflection; end-to-end agentic RL then optimizes the full long-horizon trajectory.

⚖️

Hybrid Reward Mechanism

Pointwise outcome rewards judge final image quality, while pairwise process rewards guarantee monotonic improvement across iterative reflections.

🔁

Round-Aware Resampling

Uniformly samples trajectories across interaction rounds to prevent collapse into a single mode, balancing rewriting, judgment, and reflection skills.

Overview of GenAgent. (Left) Iterative interaction between the agentic multimodal model and image generation tools. (Middle) Performance scaling of GenAgent across interaction rounds. (Right) Zero-shot generalization to image generators of varying capability.

Method

GenAgent reasons, generates, judges, and reflects in a closed agentic loop. A unified multimodal policy is trained via cold-start SFT followed by agentic GRPO with a hybrid reward.

The GenAgent pipeline. Upon receiving a user query, the multimodal policy executes a reasoning trace that interleaves thinking, tool invocation, judgment, and reflection. During agentic RL rollouts, we compute pointwise rewards from the final image and pairwise rewards based on consecutive image pairs throughout the trajectory.

Pipeline at a glance

Each rollout is a four-phase cycle that can repeat up to n_max = 3 rounds.

Phase 1

Thinking & Rewriting

The agent reasons over the user query q and emits a refined, tool-callable prompt P₁, producing a transparent reasoning trace T₁.

Phase 2

Tool Invocation

An external image generator (FLUX.1-dev, Qwen-Image, Sana 1.5, …) is invoked as a callable tool and returns the candidate image I₁.

Phase 3

Visual Judgment

The same multimodal policy evaluates I_k against q, emitting a judgment trace J_k with explicit deficiencies if criteria are not met.

Phase 4

Reflection

If unsatisfied, the agent reflects and proposes a new prompt P_k+1; otherwise it terminates with action a. Trajectories form an interleaved multimodal chain-of-thought.

Why a cold start?

Zero-shot multimodal models exhibit three critical failure modes that doom direct RL: 13.4% tool-invocation errors, 49.3% trivial rewrites (final prompt within 5 words of the user query), and only 0.36% improvement from reflection. Hint-guided SFT on 32K trajectories drives these errors to near-zero and unlocks meaningful refinement.

Pointwise + Pairwise reward

Pointwise reward (r_point ∈ {0, 0.7}) checks whether the final image satisfies every requested condition via a generative reward model. Pairwise reward (r_pair) only fires when every consecutive image pair improves — explicitly supervising the process of reflection rather than just the outcome.

Round-Aware Resampling

We oversample G′ = 12 rollouts per query and downsample to G = 8 by uniformly drawing across subsets categorized by tool-invocation count. This prevents the policy from collapsing onto either premature termination or empty looping, preserving balanced rewriting / judgment / reflection skills.

🏆 GenEval++ Leaderboard

All decoupled methods (including GenAgent) use FLUX.1-dev as the default image generation tool with n_max=2. GenAgent w/ Qwen-Image reaches 0.725, nearly matching closed-source GPT-4o (0.739). Click column headers to sort.

Model Performance on GenEval++

#	Method	Color ▲	Count ▲	Color/Count ▲	Color/Pos ▲	Pos/Count ▲	Pos/Size ▲	Multi-Count ▲	Overall ▼

WISE: Knowledge-grounded Reasoning

Beyond compositional instruction-following, GenAgent w/ Qwen-Image surpasses every open-source unified model on knowledge-grounded reasoning and approaches closed-source GPT-4o.

Method	Cultural	Time	Space	Biology	Physics	Chemistry	Overall
Diffusion
FLUX.1-dev	0.56	0.57	0.67	0.45	0.54	0.41	0.55
Qwen-Image	0.62	0.63	0.77	0.57	0.75	0.40	0.62
Unified
GPT-4o Closed	0.81	0.71	0.89	0.83	0.79	0.74	0.80
Janus Pro 7B	0.30	0.37	0.49	0.36	0.42	0.26	0.35
T2I-R1	0.56	0.55	0.63	0.54	0.55	0.30	0.54
Bagel	0.44	0.55	0.68	0.44	0.60	0.39	0.52
Bagel w/ Self-CoT	0.76	0.69	0.75	0.65	0.75	0.58	0.70
Decoupled
PromptEnhancer	0.54	0.60	0.69	0.46	0.62	0.39	0.56
T2I-Copilot	0.71	0.66	0.77	0.56	0.67	0.48	0.67
GenAgent w/o tool Ours	0.75	0.66	0.72	0.55	0.60	0.49	0.67
GenAgent (Ours)
Qwen2.5-VL-7B Base Ours	0.70	0.63	0.68	0.53	0.61	0.41	0.63
+ SFT Ours	0.68	0.61	0.68	0.60	0.65	0.47	0.64
+ RL (w/ FLUX.1-dev) Ours	0.75	0.69	0.72	0.61	0.65	0.50	0.69
+ RL (w/ Qwen-Image) Ours	0.78	0.67	0.78	0.72	0.71	0.55	0.72

Emergent Properties

Three behaviors emerge from agentic training without being explicitly supervised.

Property 1 — Cross-Tool Generalization

Although trained only with FLUX.1-dev, GenAgent transfers zero-shot to other generators with different capability levels — Sana 1.5-1.6B (smaller), Qwen-Image (stronger) — while preserving gains. Upgrading the underlying tool yields free improvements without retraining the agent.

Property 2 — Test-Time Scaling

Performance scales monotonically across interaction rounds. The first reflection delivers the largest jump; subsequent rounds tighten residual errors. Two rounds offer the best efficiency / quality trade-off in practice.

Property 3 — Task-Adaptive Reasoning

Distinct reasoning patterns — fact-checking and correction, targeted refinement and editing, and creative fulfillment with realism — emerge automatically and are selected by the agent depending on the task at hand.

Case Studies

Selected reasoning traces illustrating GenAgent's three core patterns, plus out-of-distribution generalization to visual math.

Three distinct reasoning patterns of GenAgent. (Left) Fact-Checking and Correction: detecting that a green chameleon on brown leaves contradicts "perfectly camouflaged" and revising the prompt. (Middle) Refinement and Editing: adding "mirror-polished" to capture missing stone gloss in Stonehenge. (Right-top) Creative Fulfillment with Realism: composing a plausible cow-and-scissors scene. (Right-middle) Out-of-distribution visual math: solving GSM8K problems and converting reasoning steps into visual prompts. (Right-bottom) A representative failure mode: over-reflection on ambiguous prompts.

A representative end-to-end reasoning trace. The agent rewrites the user query, invokes the image generator, judges the output against the requested conditions, and reflects with a targeted prompt revision when criteria are unmet.

🔍

Fact-Checking

Detects logical conflicts between scene attributes and intent, then rewrites the prompt to enforce consistency — e.g., matching chameleon color to its background.

🖌️

Targeted Editing

For near-correct outputs, the agent performs local edits via additive descriptors rather than full regeneration, avoiding catastrophic redrawing.

🎨

Creative Realism

For impossible compositions, the agent constructs plausibility through composition, lighting, and perspective — balancing imagination with grounding.

Citation

If you find GenAgent useful in your research, please consider citing our paper.

@inproceedings{jiang2026genagent,
  title     = {GenAgent: Scaling Text-to-Image Generation via Agentic Multimodal Reasoning},
  author    = {Jiang, Kaixun and Wang, Yuzheng and Zhou, Junjie and Li, Pandeng and
               Liu, Zhihang and Xie, Chen-Wei and Chen, Zhaoyu and Zheng, Yun and Zhang, Wenqiang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}