GenRouter: Unified Workflow Routing for Agentic Image Generation

Harold Haodong Chen; Zhiyu Hou; Wen-Jie Shu; Weilin Ruan; Yingjie Xu; Litao Guo; Ying-Cong Chen

cs.CVAug 17, 2026

GenRouter: Unified Workflow Routing for Agentic Image Generation

Authors: Harold Haodong Chen, Zhiyu Hou, Wen-Jie Shu, Weilin Ruan, Yingjie Xu, Litao Guo, Ying-Cong Chen

Organizations: HKUST(GZ) · HKUST · SUSTech · ZODA ·

Abstract

The rapid evolution of text-to-image (T2I) generation models has effectively solved the foundational challenge of raw pixel synthesis, shifting the community's focus toward fulfilling increasingly intricate user requests. While recent agentic image generation workflows enhance static inference with advanced capabilities like external knowledge retrieval and iterative reasoning, they mostly operate in isolated silos with fixed ``one-size-fits-all" topologies. This inevitably leads to severe compute-mismatch, where simple queries are forced through computationally heavy pipelines. To bridge this gap, we present GenRouter, the first unified workflow routing framework for agentic image generation. We first formulate GenCanvas, standardizing diverse agentic pipelines into a universal set of foundational primitives and executable templates. Operating over this unified space, GenRouter adaptively routes heterogeneous prompts to their optimal workflows via (i) demand profiling, (ii) experience matching, and (iii) Pareto filtering. Extensive experiments across diverse benchmarks demonstrate that GenRouter achieves superior visual alignment while reducing execution costs by over 95% and latency by 65% compared to heavyweight static pipelines. Furthermore, the system continuously self-evolves via accumulated experience, enabling robust zero-shot generalization that boosts performance and halves computational overhead.

Figures & tables

Figure 1: ( Left ) Existing systems enforce a fixed execution workflow. ( Right ) GenRouter dynamically routes prompts to their optimal workflow trajectories over the unified GenCanvas .

Figure 2: Overview of our proposed ( Left ) GenCanvas and ( Right ) GenRouter .

Cognitive Level	Workflow ( $w\in\mathcal{W}$ )	Topological Template	Representative Method
I: Semantic Alignment	$\mathtt{DirectGen}$	$g$
	$\mathtt{RewriteGen}$	$\pi_{\mathsf{rewrite}}\rightarrow g$	BeautifulPrompt [ 3 ]
II: External Grounding	$\mathtt{SearchGen}$	$\pi_{\mathsf{search}}\text{ (text)}\rightarrow\pi_{\mathsf{rewrite}}\rightarrow g$	Gen-Searcher [ 9 ]
	$\mathtt{RefGen}$	$\pi_{\mathsf{search}}\text{ (visual)}\rightarrow\pi_{\mathsf{rewrite}}\rightarrow g$	Mind-Brush [ 14 ]
III: Structural Reasoning	$\mathtt{ReasonGen}$	$\pi_{\mathsf{reason}}\rightarrow\pi_{\mathsf{rewrite}}\rightarrow g$	Self-CoT [ 8 ]
	$\mathtt{SkillGen}$	$\pi_{\mathsf{skill}}\rightarrow\pi_{\mathsf{rewrite}}\rightarrow g$	GEMS [ 15 ]

Table 1: Core workflow templates and topological paths in GenCanvas . We abstract diverse generative demands into

9

structured templates over

\Pi

. Representative methods are mapped solely for conceptual alignment, not strict implementation equivalence. Icons classify these prior works into test-time scaling ( ) or agentic workflow ( ) for image generation.

Generator	Method	WISE			DPG-Bench			GenEval2
Generator	Method	Perf. (%)	Cost ($)	Latency (h)	Perf. (%)	Cost ($)	Latency (h)	Perf. (%)	Cost ($)	Latency (h)
Qwen-Image	Original [ 30 ]	0.53	0.00	0.00	87.21	0.00	0.00	29.0	0.00	0.00
	Mind-Brush [ 14 ]	0.68	4.40	3.65	83.74	2.42	2.71	32.1	1.80	2.38
	SCOPE [ 22 ]	0.88	1.97	4.34	86.94	4.88	9.42	46.3	1.81	3.89
	GEMS [ 15 ]	0.80	38.17	11.94	85.59	60.31	15.23	70.4	40.43	8.43
	GenCanvas + GenRouter	0.88	1.57	4.16	87.39	1.41	2.52	71.6	2.03	3.59

Table 2: Performance comparison on WISE, DPG-Bench and GenEval2 benchmarks. The best and second best results are highlighted.

Generator	Method	OneIG-EN			OneIG-CN			Average
Generator	Method	Perf. (%)	Cost ($)	Latency (h)	Perf. (%)	Cost ($)	Latency (h)	Perf. (%)	Cost ($)	Latency (h)
Qwen-Image	Original [ 30 ]	0.487	0.00	0.00	0.489	0.00	0.00	53.4	0.00	0.00
	Mind-Brush [ 14 ]	0.508	3.30	4.95	0.516	4.34	7.75	57.2	3.25	4.29
	SCOPE [ 22 ]	0.521	4.30	8.87	0.504	4.92	10.39	64.7	3.58	7.38
	GEMS [ 15 ]	0.542	73.70	16.21	0.532	86.90	16.27	68.7	59.70	13.62
	GenCanvas + GenRouter	0.553	5.50	5.92	0.541	4.34	7.23	71.3	2.97	4.68

Table 3: Performance comparison on OneIG-Bench benchmarks. “Average" includes five benchmarks’ results in both Tables 2 and 3 .

Figure 3: Qualitative comparison. Left side shows cost (scaled by $0.01$ ); right side shows latency.

Figure 4: ( Left ) Routing distribution of GenRouter ; ( Middle ) Utility comparison between isolated GenCanvas workflows; ( Right ) Zero-shot generalization on DPG-Bench with WISE experience.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Workflow demonstration of GenCanvas & GenRouter .

Figure 7: Workflow demonstration of GenCanvas & GenRouter .

Figure 8: Case study of GenCanvas & GenRouter .

Figure 9: Case study of GenCanvas & GenRouter .

Info.	WISE	DPG	OneIG-EN	OneIG-CN	LongText-EN	LongText-CN	SpatialGenEval	ArtiMuse	GenEval2
Quantity	100	50	50	50	50	50	50	50	50
Score Weight	1.12	1.21	1.44	1.43	1.03	1.05	1.63	1.69	1.59

Appendix

Table 4: Composition of the mixed test set.

Figure 10: Analysis of bounded space of GenCanvas .

Figure 11: Analysis of components of GenRouter .

Utility	Performance $\uparrow$	Cost ( $)$ \downarrow$	Latency (h) $\downarrow$
$\lambda_{c}=1.0$	73.56	1.56	1.98
$\lambda_{c}=5.0$	73.52	1.37	1.76
$\lambda_{c}=10.0$	73.00	1.36	1.87
$\lambda_{l}=0.0003$	73.26	1.50	1.85
$\lambda_{l}=0.0006$	73.52	1.37	1.76
$\lambda_{l}=0.0010$	73.47	1.37	1.74

Appendix

Table 5: Sensitivity analysis of GenRouter + on the mixed test set.

Method	Performance $\uparrow$	Cost ( $)$ \downarrow$	Latency (h) $\downarrow$
Original	62.51	0	0
+ $\mathcal{M}_{\text{traj}}$ only	72.18	1.29	1.71
+ $\mathcal{M}_{\text{route}}$ only	72.15	1.13	1.57
+ GenRouter	73.52	1.37	1.76

Appendix

Table 6: Memory analysis of GenRouter + on the mixed test set.

Figure 12: More results demonstration of GenRouter within GenCanvas .

Figure 13: More results demonstration of GenRouter within GenCanvas .

Agentic Image Generation

Agentic image generation uses reasoning, tools, search, and iterative feedback to create and refine images with greater control.

See all 43 papers

OctoT2I: A Self-Evolving Agentic Text-to-Image Router
Jun 1, 2026Xu Jiang, Bin Chen, Gehui Li +3T2I Generation Agentic Image Generation
School of Electronic and Computer Engineering, Peking University · Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University
Generation Navigator: A State-Aware Agentic Framework for Image Generation
May 18, 2026Jinming Liu, Ruoyu Feng, Yuqi Wang +2T2I Generation Agentic Image Generation
Shanghai Jiao Tong University · Independent · Eastern Institute of Technology, Ningbo, China
Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation
Jun 25, 2026Zekai Zhang, Jiahao Li, Jie Zhang +18Conditional Image Generation T2I Generation

Jun 1, 2026J/K move · Enter open · S save

OctoT2I: A Self-Evolving Agentic Text-to-Image Router

Xu Jiang, Bin Chen, Gehui Li, Yule Duan+2

School of Electronic and Computer Engineering, Peking University · Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University

T2I Generation Agentic Image Generation Compositional T2I Generation

Open paper AskPDF arXiv

The explosive growth of Text-to-Image (T2I) models, from large-scale versions to lightweight, real-time ones, now faces diminishing marginal returns from single-model scaling. Agentic T2I methods emerged to alleviate this bottleneck by using multiple models. However, existing agentic T2I methods suffer from three key challenges: reliance on expensive handcrafted priors or human annotations, rigid single-path decision mechanisms, and a neglect of inference efficiency. To address these challenges, we introduce OctoT2I, a novel agentic framework that reformulates the T2I task as a joint optimization of generation quality and inference efficiency. OctoT2I implements a stateful, multi-round routing strategy that adaptively selects the most suitable tool based on its knowledge and memory. This strategy is enabled by a knowledge base built from scratch by our novel Self-Evolving Mechanism. This mechanism, which requires no human supervision, first autonomously defines foundational Conceptual Dimensions (eg, style, color, count) and then intelligently explores their combinations via an iterative" Propose--Solve--Evaluate--Learn"(PSEL) loop. The PSEL loop efficiently discovers each tool's capability frontier, driving continuous improvement without external guidance. Extensive experiments demonstrate that OctoT2I achieves competitive performance (0.96) on GenEval while delivering a 90.3% inference speedup and a 56.6% energy-efficiency gain over the leading baseline (Flow-GRPO), striking an exceptional balance between performance and efficiency. Code and models will be made available.

Abstract

Agentic Image Generation

Explore similar work

OctoT2I: A Self-Evolving Agentic Text-to-Image Router

Generation Navigator: A State-Aware Agentic Framework for Image Generation

Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation