Infographics no longer force a choice: global redraw and local text repair split routes, letting open-weight generators beat closed models

Infographic generation can now choose the scope of a revision according to the size of the error without changing generator weights: redraw when the overall composition fails, edit only one block when a few labels or formulas fail. Earlier agentic methods either rewrite the prompt once or rely on full-image regeneration; the former struggles to enforce fine-grained constraints, and the latter perturbs regions that already passed verification. On BizGenEval, Stride raises the frozen ERNIE-Image and Infographic-V2 totals from 28.0 and 41.1 to 51.1 and 69.5, which is 23.4 and 26.3 points above the strongest agentic baseline, Gen-Searcher, on the same generators.

Before each revision, the system decides which constraint failed, locates where it failed in the image, and only then picks which subagent acts. Judgment and grounding are split between two multimodal models: a verifier returns pass or fail for each atomic rubric item, and a perceptor returns region-level descriptions covering object locations, rendered text, and attribute-to-region bindings. In the BizGenEval ablation, adding the perceptor to the lead-editor, composer and retoucher configuration lifts the three-iteration hard score from 48.9 to 51.1; the lead editor's routing also tracks error type, sending about ninety percent of layout decisions to the composer and about eighty percent of text decisions to the retoucher.

A candidate image replaces the current champion only when the verifier score strictly improves; otherwise the incumbent stays, so repeated rounds do not damage parts that are already correct. Random replacement and forced replacement both trail clearly at the same candidate budget, so the gain is not explained by extra sampling alone. On BizGenEval with four candidates, forced replacement averages 45.4 on the hard track while incumbent retention reaches 51.1; raising the candidate count from one to four moves the evolution strategy from 45.6 to 51.1.

The demonstrated scope is infographic and knowledge-constrained image generation: BizGenEval covers slides, webpages, posters, charts and scientific figures, IGenBench covers 30 infographic types, and WISE covers cultural, spatio-temporal and natural-science prompts described indirectly. The intended reader is an engineer working on text-to-image reliability and agentic refinement; the method acts on frozen generators and needs no retraining. The source proposes extending local refinement from text and formulas to structured visual elements such as chart geometry and spatial relationships.

The source states that refinement still depends on the reliability of rubric verification and spatial grounding, and that monotonic admission applies to the internal verifier score and does not guarantee improvement under independent evaluation; diagnostics show the verifier sometimes accepts inferior updates or rejects beneficial ones. The retoucher currently handles text and formula replacement within localized regions, while changes to chart geometry and spatial relationships still rely on the composer's global regeneration. This report is based on the full text and did not independently recheck the benchmark or human-evaluation raw data.