Stride: Scale-Aware Multi-Agent Evolutionary Refinement for Infographic Generation

Yan Zhou; Yang Yang

cs.CVOct 9, 2026

Stride: Scale-Aware Multi-Agent Evolutionary Refinement for Infographic Generation

Authors: Yan Zhou, Yang Yang

Organizations: School of Artificial Intelligence Shanghai Jiao Tong University

Abstract

Infographic generation challenges text-to-image models because a single image must satisfy numerous visual, textual, and semantic constraints. Iterative agentic refinement with a multimodal verifier is a natural remedy, but existing methods face a scale mismatch: local edits cannot correct global composition errors, while full regeneration may destroy already-correct regions. We introduce Stride, a constraint-grounded multi-agent framework that treats refinement over any frozen generator as a history-aware evolution strategy. Stride separates constraint evaluation - whether each requirement is satisfied - from visual grounding - where and how it is violated. It then routes global revisions to a composer subagent and local corrections to a retoucher subagent. Each subagent conducts internal planning and generates multiple candidates, while Stride accepts a candidate only when it improves the verifier score, reducing regression across iterations. Because stochastic generators and imperfect verifiers make any single step unreliable, it is this combination of scale-aware routing, per-step redundancy, and monotonic acceptance that yields stable improvement across iterations. With fully open-weight generator-verifier systems, Stride outperforms proprietary systems including Nano-Banana-2.0 and Seedream-5.0-Pro on BizGenEval, matches GPT-Image-2.0 and surpasses Seedream-5.0-Pro on IGenBench, and achieves a new state-of-the-art score on WISE, while substantially outperforming existing agentic methods across all evaluated benchmarks.

Figures & tables

Figure 1: Overview of Stride . From $K$ initial samples, Stride selects a champion and refines it for $t$ rounds. Each round performs evolutionary refinement and agent state update through dual-route evaluation: the verifier scores candidate images against the rubrics, while the perceptor provides visual grounding of the observed failures. The lead editor then analyzes the agent state and refinement trajectory to route the champion to either the composer for global redraw or the retoucher for local editing. The generated candidates are then re-evaluated, and the champion is updated only when a candidate improves the verifier score. Repeating this process yields the final champion.

Model	Venue	Params	Score				Overall
Model	Venue	Params	Layout	Attribute	Text	Knowledge	Average
Proprietary models
Nano-Banana-Pro	25@Google	–	72.2	65.6	86.4	82.6	76.7
Wan2.6-T2I	26@Alibaba	–	46.4	16.6	12.6	12.2	21.9
Seedream-5.0	26@ByteDance	–	67.6	42.4	43.4	41.8	48.8
Seedream-5.0-Pro	26@ByteDance	–	75.5	74.1	63.2	49.8	65.7

Table 1: Quantitative comparison on BizGenEval. Each entry reports the hard score. The proprietary-model results are reproduced unchanged from the official BizGenEval.

Model	Venue	Params	Question Type Q-ACC										Overall
Model	Venue	Params	Comp.	Enc.	Order	Marks	Anno.	Axes	Leg.	Chart	Title	Deco.	Q-ACC	I-ACC
Proprietary models
Image-01	25@MiniMax	–	0.01	0.05	0.04	0.10	0.10	0.14	0.03	0.22	0.14	0.47	0.13	0.00
P-Image	25@PrunaAI	–	0.08	0.15	0.19	0.27	0.36	0.28	0.54	0.43	0.58	0.68	0.34	0.00
Nano-Banana	25@Google	–	0.18	0.31	0.27	0.44	0.54	0.57	0.52	0.60	0.65	0.81	0.48	0.02
Nano-Banana-Pro	25@Google	–	0.84	0.86	0.90	0.87	0.93	0.93	0.96	0.92	0.98	0.94	0.90	0.49

Table 2: Quantitative comparison on IGenBench. The proprietary-model results are reproduced unchanged from the official IGenBench.

Figure 2: Qualitative comparison with existing agentic methods. Red boxes highlight errors.

$\#$	Components					BizGenEval									Avg.
	Lead	Comp.	Ret.	Perc.	Iters.	Slides	Webpage	Poster	Chart	Sci. Fig.	Layout	Attr.	Text	Know.	Score
(1)	Best-of-1				1 iter	32.8	41.5	35.6	10.5	19.7	51.8	31.8	28.3	0.1	28.0
(2)	Best-of-6				1 iter	39.0	48.5	40.2	19.0	26.8	60.6	41.4	36.4	0.4	34.7
(3)	✓	✓	–	–	3 iters	48.5	53.2	44.8	21.2	38.0	63.8	45.4	40.6	14.8	41.1
(4)	✓	–	✓	–	3 iters	57.5	60.5	52.0	38.8	42.5	64.4	41.4	60.4	34.8	50.2
(5)	✓	✓	✓	–	3 iters	57.0	57.0	54.5	34.8	41.5	64.2	49.4	51.0	31.2	48.9

Table 3: Stride ablation study with ERNIE-Image as the base generator. We compare sampling-only baselines, execution routes, perception grounding, and refinement iterations on BizGenEval.

Table 6

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Scale	Coverage	Final eval. units	Metrics
BizGenEval	400 prompts	5 domains $\times$ 4 capabilities = 20 tasks	8,000 checklist judgments	Score, Accuracy
IGenBench	600 cases	30 infographic types in 6 categories	5,259 QA judgments	Q-ACC, I-ACC
WISE	1,000 prompts	3 domains, 25 subdomains, 6 report categories	3,000 dimension scores	WiScore

Appendix

Table 6: Quantitative overview of the public benchmarks used in our evaluation. Final reporting follows each benchmark’s official evaluator and metrics.

Method	Venue	A / G / V	WebSearch	Rubric Feedback	Global Redraw	Local Edit	Iterative Refine.	Champion Admit
ReflectionFlow	25@ICCV	– / O / O	$\times$	$\times$	$\surd$	$\times$	$\surd$	$\times$
T2I-Copilot	25@ICCV	P / O / O	$\times$	$\times$	$\surd$	$\times$	$\surd$	$\times$
Ag3	25@ICML	P / P / P	$\times$	$\times$	$\surd$	$\times$	$\surd$	$\times$
Gen-Searcher	26@Arxiv	O / P / –	$\surd$	$\times$	$\times$	$\times$	$\times$	$\times$
GenEvolve	26@Arxiv	O / P / –	$\surd$	$\times$	$\times$	$\times$	$\times$	$\times$
Mind-Brush	26@SIGAsia	P / P / –	$\surd$	$\times$	$\times$	$\times$	$\times$	$\times$

Appendix

Table 7: Comparison of agentic methods on the mechanisms required for infographic refinement in the inference time. A, G and V for the agent backbone, base generator, and verifier, with O, P, and – denoting open-source or open-weight, proprietary, and not applicable.

Model	Venue	Params	Visual Domains					Capabilities				Average
Model	Venue	Params	Slides	Webpage	Poster	Chart	Sci. Figure	Layout	Attribute	Text	Knowledge	Average
Proprietary models
Nano-Banana-Pro	25@Google	–	82.2 (94.8)	77.5 (96.5)	76.5 (94.8)	73.0 (92.2)	74.2 (90.0)	72.2 (91.2)	65.6 (92.2)	86.4 (95.0)	82.6 (96.2)	76.7 (93.7)
Wan2.6-T2I	26@Alibaba	–	27.1 (56.5)	25.5 (67.0)	27.5 (62.5)	17.2 (48.8)	12.5 (58.8)	46.4 (80.6)	16.6 (60.6)	12.6 (52.6)	12.2 (41.0)	21.9 (58.7)
Seedream-5.0	26@ByteDance	–	54.5 (80.8)	47.0 (80.8)	50.7 (77.0)	46.0 (76.2)	45.8 (81.5)	67.6 (89.0)	42.4 (77.2)	43.4 (75.6)	41.8 (75.2)	48.8 (79.2)
Seedream-5.0-Pro	26@ByteDance	–	–	–	–	–	–	75.5 (87.6)	74.1 (87.3)	63.2 (81.1)	49.8 (75.3)	65.7 (82.8)

Appendix

Table 8: Extended quantitative comparison on BizGenEval. Each entry reports the hard score followed by the easy score in parentheses. Rows are ordered by Venue and, within each year, by hard Average. Params denotes the parameters of the base generator. The proprietary-model results are reproduced unchanged from the official table of BizGenEval. Missing entries are denoted by “–”.

$\#$	Components					BizGenEval									Average	WISE
	Lead	Comp.	Ret.	Perc.	Iters.	Slides	Webpage	Poster	Chart	Sci. Fig.	Layout	Attr.	Text	Know.	Score	Overall
(1)	Best-of-1				1 iter	32.8 (57.0)	41.5 (65.6)	35.6 (62.8)	10.5 (35.3)	19.7 (55.4)	51.8 (78.7)	31.8 (73.9)	28.3 (65.1)	0.1 (3.1)	28.0 (55.2)	0.58
(2)	Best-of-6				1 iter	39.0 (58.2)	48.5 (69.0)	40.2 (64.8)	19.0 (44.8)	26.8 (58.0)	60.6 (81.0)	41.4 (80.4)	36.4 (70.6)	0.4 (3.8)	34.7 (59.0)	–
(3)	✓	✓	–	–	1 iter	44.5 (65.0)	53.5 (71.2)	45.5 (64.5)	21.2 (46.5)	37.5 (69.0)	63.8 (81.6)	43.0 (80.4)	39.4 (69.2)	15.6 (21.8)	40.4 (63.2)	0.89
(4)	✓	✓	–	–	2 iters	44.5 (64.0)	54.2 (70.5)	44.5 (66.0)	22.2 (48.8)	38.7 (69.8)	64.4 (81.6)	43.2 (80.6)	40.8 (68.8)	15.0 (24.2)	40.8 (63.8)	–
(5)	✓	✓	–	–	3 iters	48.5 (66.5)	53.2 (69.5)	44.8 (65.0)	21.2 (49.0)	38.0 (68.8)	63.8 (80.6)	45.4 (78.4)	40.6 (70.0)	14.8 (26.0)	41.1 (63.7)	–

Appendix

Table 9: Extended Stride ablation study on BizGenEval and WISE benchmarks.

Setup	Dimension	Iter 1	Iter 2	Iter 3	Overall
ERNIE-Image	Layout	97.2 / 2.8 ( $72$ )	92.7 / 7.3 ( $55$ )	89.6 / 10.4 ( $48$ )	93.7 / 6.3 ( $175$ )
	Attribute	82.7 / 17.3 ( $81$ )	77.3 / 22.7 ( $75$ )	81.9 / 18.1 ( $72$ )	80.7 / 19.3 ( $228$ )
	Text	9.2 / 90.8 ( $87$ )	17.6 / 82.4 ( $68$ )	23.7 / 76.3 ( $59$ )	15.9 / 84.1 ( $214$ )
	Knowledge	55.1 / 44.9 ( $98$ )	53.4 / 46.6 ( $88$ )	42.2 / 57.8 ( $83$ )	50.6 / 49.4 ( $269$ )
	Total	58.9 / 41.1 ( $338$ )	58.7 / 41.3 ( $286$ )	57.6 / 42.4 ( $262$ )	58.5 / 41.5 ( $886$ )
Infographic-V2	Layout	94.4 / 5.6 ( $72$ )	93.4 / 6.6 ( $61$ )	96.2 / 3.8 ( $52$ )	94.6 / 5.4 ( $185$ )

Appendix

Table 10: Per-iteration scale-aware sub-agent preference of the lead editor, reported as composer / retoucher percentages on the BizGenEval with two generators. The number of rounds is given in parentheses.

Model	Venue	Params	Visual Domains					Capabilities				Average
Model	Venue	Params	Slides	Webpage	Poster	Chart	Sci. Figure	Layout	Attribute	Text	Knowledge	Average
Proprietary models
Nano-Banana-Pro	25@Google	–	82.2 (94.8)	77.5 (96.5)	76.5 (94.8)	73.0 (92.2)	74.2 (90.0)	72.2 (91.2)	65.6 (92.2)	86.4 (95.0)	82.6 (96.2)	76.7 (93.7)
Wan2.6-T2I	26@Alibaba	–	27.1 (56.5)	25.5 (67.0)	27.5 (62.5)	17.2 (48.8)	12.5 (58.8)	46.4 (80.6)	16.6 (60.6)	12.6 (52.6)	12.2 (41.0)	21.9 (58.7)
Seedream-5.0	26@ByteDance	–	54.5 (80.8)	47.0 (80.8)	50.7 (77.0)	46.0 (76.2)	45.8 (81.5)	67.6 (89.0)	42.4 (77.2)	43.4 (75.6)	41.8 (75.2)	48.8 (79.2)
Seedream-5.0-Pro	26@ByteDance	–	–	–	–	–	–	75.5 (87.6)	74.1 (87.3)	63.2 (81.1)	49.8 (75.3)	65.7 (82.8)

Appendix

Table 11: Extended quantitative comparison on BizGenEval with oracle experiments. Each entry reports the hard score followed by the easy score in parentheses. Rows are ordered by Venue and, within each year, by hard average. Params denotes the parameters of the base generator. The proprietary-model results are reproduced unchanged from the official table of BizGenEval benchmark. Missing entries are denoted by “–”.

Model	Venue	Params	Question Type Q-ACC										Overall
Model	Venue	Params	Comp.	Enc.	Order	Marks	Anno.	Axes	Leg.	Chart	Title	Deco.	Q-ACC	I-ACC
Proprietary models
Image-01	25@MiniMax	–	0.01	0.05	0.04	0.10	0.10	0.14	0.03	0.22	0.14	0.47	0.13	0.00
P-Image	25@PrunaAI	–	0.08	0.15	0.19	0.27	0.36	0.28	0.54	0.43	0.58	0.68	0.34	0.00
Nano-Banana	25@Google	–	0.18	0.31	0.27	0.44	0.54	0.57	0.52	0.60	0.65	0.81	0.48	0.02
Nano-Banana-Pro	25@Google	–	0.84	0.86	0.90	0.87	0.93	0.93	0.96	0.92	0.98	0.94	0.90	0.49

Appendix

Table 12: Extended quantitative comparison on IGenBench with oracle experiments. Params denotes the parameters of the base generator. The proprietary-model results are reproduced unchanged from the official IGenBench.

Model	Venue	Params	Cult.	Time	Space	Bio.	Phys.	Chem.	Overall
Proprietary models
Seedream 4.0	25@ByteDance	–	0.78	0.73	0.85	0.79	0.84	0.67	0.78
GPT-4o / GPT-Image-1	25@OpenAI	–	0.81	0.71	0.89	0.83	0.79	0.74	0.80
Nano-Banana-Pro	25@Google	–	0.89	0.80	0.89	0.88	0.86	0.85	0.87
Open-weight models
FLUX.1-dev	24@BFL	12B	0.48	0.58	0.62	0.42	0.51	0.35	0.50

Appendix

Table 13: Quantitative comparison on WISE. We report WiScore across six knowledge domains and the overall score. Best and second-best results are shown in bold and underline, respectively. The proprietary-model results are reproduced unchanged from the official WISE.

Benchmark	Generator	Prompt pairs	Images	Human rubric
BizGenEval	ERNIE-Image	25	50	$25\times 10$ Hard questions
BizGenEval	Infographic-V2	25	50	$25\times 10$ Hard questions
IGenBench	ERNIE-Image	25	50	All released atomic questions
IGenBench	Infographic-V2	25	50	All released atomic questions
Total	–	100	200	–

Appendix

Table 14: Sampling protocol for the human–MLLM evaluation study. For each prompt, we retain paired outputs from the frozen generator and Stride under the same generator and prompt.

Figure 3: Best-of- $K$ selection on BizGenEval over 400 items on the hard track. (a) Each verifier scores its own pick. (b) The same picks rescored by Gemini-3-Flash-Preview. Hard average values are mean per-item checklist accuracy on the hard track, whcih differ from error score (Appendix B ).

Selector	Self	Gemini	Capture	Leniency	Inflation
Gemini (Oracle)	57.7	57.7	100%	$+0.0$	$+0.0$
Qwen3.5-397B-A17B	60.7	51.6	40.2%	$+1.7$	$+9.1$
Qwen3.7-Plus	60.9	50.5	29.2%	$+2.2$	$+10.4$
Kimi-K2.7-HS	69.3	50.5	29.4%	$+7.8$	$+18.8$
Kimi-K2.5	66.6	50.4	27.9%	$+7.3$	$+16.2$
GPT-5.5	56.0	50.1	25.0%	$-1.4$	$+5.9$

Appendix

Table 15: Best-of-

6

selection on BizGenEval over 400 items, reported as mean checklist accuracy on the hard track. Self is the verifier’s score of its own pick. Gemini is that pick rescored by Gemini. Capture is the fraction of the Gemini gain from

47.5

57.7

that each verifier realizes. Leniency is Self minus Gemini at

k=1

, where the images coincide. Inflation is Self minus Gemini at

k=6

Figure 4: Qualitative comparison of perceptor outputs on generated infographics. Different perception backbones identify visible text blocks, localize them with bounding boxes, and assess whether the rendered text is garbled. Green boxes indicate regions that the perceptor identifies as correct, whereas red boxes indicate regions it identifies as incorrect. The comparison highlights the trade-off between geometric localization and contextual text understanding that motivates the use of an MLLM-based perceptor in Stride .

Model	A-R	A-P	A-F1	mean IoU	NED-sim	exact	Anchor med. / P90	Med. latency
PP-OCRv5	0.881	0.703	0.764	0.729	0.899	0.590	2.2 / 6.6	44s
Kimi-K3	0.738	0.699	0.713	0.730	0.949	0.721	4.6 / 16.3	49s
Qwen3.5-397B-A17B	0.511	0.486	0.495	0.749	0.957	0.751	6.0 / 46.3	71s
Qwen3.8-Max	0.388	0.360	0.370	0.670	0.924	0.728	13.4 / 148.4	259s
LocateAnything-3B	0.345	0.535	0.322	0.777	–	–	–	6s
Kimi-K2.7-Code-HS	0.318	0.292	0.300	0.654	0.896	0.698	16.0 / 98.6	29s

Appendix

Table 16: Perception-backbone agreement with Gemini-3-Flash-Preview on 20 stratified generated infographics under a box-first protocol. PP-OCRv5 is reported as a cross-check/reference OCR baseline. Metrics measure agreement with the proxy reference, not absolute accuracy. Higher is better for A-R, A-P, A-F1, mean IoU, NED-sim, and exact; lower is better for anchor offset and latency.

Non-oracle strategies
Strategy	Content Domain					Capability				Average
Strategy	Slides	Webpage	Poster	Chart	Sci. Fig.	Layout	Attr.	Text	Know.	Average
Random selection	45.0 (74.5)	59.8 (77.2)	47.5 ( 73.8 )	21.3 (50.0)	28.5 (67.0)	54.6 (78.2)	40.0 (77.6)	46.4 ( 77.0 )	20.6 (41.2)	40.4 (68.5)
$(1,1)$ forced replacement	47.5 (71.8)	55.0 (77.2)	44.2 (65.0)	22.0 (43.5)	28.0 (60.8)	56.0 (75.2)	37.8 (73.8)	42.2 (72.8)	21.4 (32.8)	39.4 (63.7)
$(1+1)$ -evolution strategy	53.0 (75.0)	60.2 (81.2)	50.3 (72.5)	29.8 ( 53.5 )	34.5 (65.5)	62.6 ( 80.0 )	43.6 ( 81.0 )	49.2 (75.8)	26.8 (41.4)	45.6 ( 69.5 )
$(1,4)$ forced replacement	56.0 ( 76.2 )	62.0 ( 81.5 )	49.3 (72.2)	27.8 (49.8)	32.0 ( 67.2 )	59.2 (77.2)	42.4 (77.8)	50.4 ( 79.8 )	29.6 ( 42.8 )	45.4 (69.4)

Appendix

Table 17: Comparison of champion-evolution strategies on BizGenEval. Each entry reports the official error score as hard (easy). All non-oracle variants use the same refinement framework and differ only in how new champion candidates are selected and admitted. The

(1,1)

and

(1,4)

variants denote forced replacement, whereas the

(1+1)

and

(1+4)

variants retain the incumbent champion unless a better candidate is found. Oracle admission keeps Qwen3.5-397B-A17B within-round candidate selection unchanged but uses Gemini-3-Flash-Preview for admission; Full oracle uses Gemini-3-Flash-Preview for both candidate selection and admission.

Figure 5: More qualitative comparison with existing agentic methods on the infographic task.

Figure 6: More qualitative results of Stride .

Figure 7: Qualitative comparison with existing agentic methods on general WISE task.

Figure 8: Case study 1: Localized text repair with Stride . Guided by visual grounding, the lead editor selects local editing to correct a garbled text block while preserving the surrounding content and layout. Green and red boxes mark regions identified as correct and erroneous, respectively.

Figure 9: Case study 2: Multiple text corrections with Stride . Despite low initial verification scores, the lead editor identifies localized label errors and selects local editing to correct phonetic transcriptions and a Chinese character while preserving the overall layout. Green and red boxes mark regions identified as correct and erroneous, respectively.

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

Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation
Oct 12, 2026Zixu Wang, Xiaoran Zhao, Yahui Liu +8Compositional Image Generation Agentic Image Generation
NUDT · PDL Lab · Xiaohongshu +4
GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation
May 20, 2026Sixiang Chen, Zhaohu Xing, Tian Ye +7Agentic Image Generation LLM Agent Self-Improvement
The Hong Kong University of Science and Technology (Guangzhou) · Meituan · The Hong Kong University of Science and Technology +1
RAISE: Requirement-Adaptive Evolutionary Refinement for Training-Free Text-to-Image Alignment
Feb 28, 2026Liyao Jiang, Ruichen Chen, Chao Gao +1T2I Generation Text-Image Alignment
Department of ECE, University of Alberta, Canada · Huawei Technologies, Canada

Oct 12, 2026J/K move · Enter open · S save

Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation

Zixu Wang, Xiaoran Zhao, Yahui Liu, Feng-ao Wang+7

NUDT · PDL Lab · Xiaohongshu+4

Compositional Image Generation Agentic Image Generation Compositional T2I Generation RL for Image Generation

Open paper AskPDF arXiv

Text-to-image models have made remarkable progress in visual fidelity and instruction following, yet complex compositional generation remains challenging. Existing agentic approaches address this limitation through iterative generation and editing, adapting their actions based on interaction history. However, they typically fix the continuation state (i.e., the image from which refinement proceeds) to the latest image. Since iterative editing is not necessarily monotonic, the agent may continue refining from the latest image even when it has already degraded. In this work, we introduce GenRevisit, an agentic framework that extends action selection to joint continuation-state and action selection. The agent can not only use interaction history to adjust editing instructions but also select an image candidate to edit, jointly deciding where to continue from and how to refine. We train GenRevisit with SFT followed by agentic reinforcement learning. To address the coarse credit assignment of final outcome reward in GRPO, which assigns the same trajectory-level advantage to all actions, we further propose Normalized Action Credit Assignment (NACA) to redistribute positive advantage according to each action's contribution to reducing the remaining objective. On GenEval2, GenRevisit achieves GM scores of 84.3 and 74.4 with Qwen-Image-2512 and Z-Image-Turbo, surpassing the strongest baselines using the same generators by 12.7 and 10.9 points respectively. GenRevisit achieves state-of-the-art performance on GenEval, GenEval++, and GenEval2 among baselines.

Abstract

Agentic Image Generation

Explore similar work

Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation

GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

RAISE: Requirement-Adaptive Evolutionary Refinement for Training-Free Text-to-Image Alignment