AesCode Infographics Style: What the aesthetic dimension measures and how to interpret its score
Image generated by the Inferama Editorial Team
01

Scope and evaluated unit

The Style dimension in the AesCode benchmark does not measure data accuracy or code functionality, but the perceived aesthetic quality of the generated visual artifact. It sits within the aggregated Visual metric, which also includes the Content dimension (presence of required elements) and Layout (spatial organization and relations).

This separation is fundamental: an infographic can score high on Style due to its typography, color usage, and visual hierarchy, while failing on the accuracy of the presented data. The generation prompt remains the source of truth for content, but the Style evaluation focuses on how that content is presented on the canvas, not on whether the data is exact.

The three Visual dimensions answer different questions about the same rendered artifact. Content checks whether required elements appear, Layout checks how those elements are arranged and related, and Style assesses their visual treatment. Because each dimension has its own sample-specific yes/no criteria, a result in one dimension should not be treated as a substitute for the others.

The generation prompt establishes what the infographic is meant to contain, but Style evaluates the presentation of that content on the canvas. A polished visual treatment can therefore coexist with missing elements or inaccurate values; the Style result alone cannot establish that the artifact faithfully represents the prompt or its data.

Breakdown of the Visual metric in AesCode

The Visual score is the average of three independent dimensions evaluated by a VLM.

DimensionWhat it evaluatesEvaluation method
ContentPresence and completeness of elements required by the promptSample-specific yes/no questions
LayoutOrganization, spatial relations, and collisionsSample-specific yes/no questions
StyleHierarchy, typography, color, and component finishSample-specific yes/no questions
02

Methodology reconstruction

The Style score is obtained through a VLM judge (Vision-Language Model) that answers a list of boolean questions derived from the prompt requirements. These questions form part of the Visual Graph Rubric (VGR), a set of sample-specific criteria generated from the target design and frozen before evaluation.

Unlike deterministic metrics such as Text or Boundary, which are calculated through programmatic verification of code and rendering, Style depends on the visual perception of the evaluating model. The VGR ensures that questions are linked to specific elements of the design graph, avoiding ambiguous holistic judgments. The final score is the pass rate over these questions.

The VGR rubric connects each Style question to a particular element or relation in the target design. This makes the evaluation more specific than a single overall impression, while leaving the judgment itself dependent on a VLM’s reading of the rendered image. The rubric is fixed before scoring, so the criteria are not changed in response to an individual output.

The pass rate expresses the share of Style questions the judge answers affirmatively. It is therefore a result over the rubric’s defined questions, not a direct measurement of every possible aesthetic quality. Text and Boundary use programmatic checks, whereas Style uses visual judgment; these methods capture different properties and should not be treated as interchangeable.

Evaluation flow for the Style dimension

Step-by-step process to obtain the Style score in an AesCode run.

  1. 01Generation of HTML/CSS code from the prompt and reference image.
  2. 02Rendering of the code in a browser to obtain the final image.
  3. 03Extraction of the sample-specific VGR rubric (yes/no questions linked to the design).
  4. 04Evaluation of the rendering by the VLM judge answering the Style questions.
  5. 05Calculation of the pass rate over the Style questions.
03

What a score allows concluding

A high Style score indicates that the artifact meets the aesthetic criteria defined in the rubric, such as chromatic coherence, typographic legibility, and the absence of unintended overlaps. It does not guarantee that the data is correct, that the code is efficient, or that the infographic is useful for its communicative purpose.

It is crucial to distinguish between aesthetic quality and factual correctness. A model can produce a visually appealing design (high Style) but with errors in chart values (low Rule). Therefore, the Style score must always be interpreted together with the Rule metrics (Text, Boundary, Chart) to obtain a complete view of the artifact's quality.

A high score means that the artifact satisfies a larger share of the rubric’s specified aesthetic checks, including criteria such as legibility and visual coherence. It does not certify factual accuracy or communicative usefulness: those conclusions require attention to other properties of the artifact, not just its appearance.

Reading Style alongside Text, Boundary, and Chart helps separate visual presentation from the correctness of displayed text, boundaries, and chart values. A strong Style result paired with a weak Rule result can indicate an attractive presentation whose data or other evaluated properties still fail. The combined metrics provide a broader assessment, but the Style score remains limited to its own rubric.

04

Execution conditions and comparability

For two Style scores to be comparable, the following conditions must remain constant: the generator model (AesCode-8B or AesCode-32B), the generation prompt, the reference image (if used), the renderer (browser and version), and the sample-specific VGR rubric.

The AesCode benchmark uses 300 infographic prompts, with three generations per prompt and no selection of the best output. This means the reported score is an average of multiple runs, which reduces random variance but does not eliminate sensitivity to rendering conditions and the VLM judge's configuration.

Comparability depends on holding the evaluation setup steady, not merely on using the same benchmark name. Changes to the generator, prompt, reference image, browser rendering, or sample-specific rubric can alter what is produced or what the judge is asked to assess. A score difference under changed conditions cannot be attributed to the model alone without accounting for those changes.

Three generations for each of the 300 prompts contribute to the reported average, and selecting no best output avoids basing that average on a chosen example. Repeating generations can reduce the influence of random variation in the reported result, but it does not remove differences caused by rendering conditions or the judge’s configuration. The average should therefore be read as a summary of those runs under the stated setup.

Critical variables for Style comparability

Factors that must be controlled when comparing Style scores between different models or configurations.

VariableDescriptionImpact on Style
Generator modelAesCode-8B vs AesCode-32BHigh: greater design capacity
Reference imagePresence or absence of visual referenceHigh: guides style and distribution
RendererBrowser and Chromium versionMedium: affects rendering fidelity
VGR rubricSample-specific questionsHigh: defines evaluation criteria
05

Measurement risks and limits

The Style metric is subject to the biases of the VLM judge, which may have its own aesthetic preferences or difficulties evaluating certain design types. Additionally, the VGR rubric, although sample-specific, may not capture all dimensions of aesthetic quality, such as semantic coherence or usability.

Another risk is score optimization without improving communicative utility. A model could learn to produce designs that satisfy the rubric's criteria (e.g., using specific colors or common typographies) without improving the infographic's clarity or effectiveness. Therefore, the Style score must be interpreted with caution and complemented with human evaluations or usability metrics.

The VLM judge and the VGR rubric impose separate limits. The judge may favor some visual treatments or find some designs harder to assess, while the rubric may omit qualities that matter to people, including semantic coherence and usability. A sample-specific checklist narrows the evaluation criteria but does not make them exhaustive or universal.

Optimizing for the listed checks may improve the pass rate without improving how clearly or effectively the infographic communicates. A coherent palette and legible type can satisfy aesthetic criteria even when chart values are wrong or the overall graphic remains confusing. Human evaluation or usability measures can address aspects that the Style checklist does not capture, but they do not change what the Style score itself measures.

06

Reading and reproduction protocol

To correctly reproduce or interpret a Style score, it is necessary to record the model version, the exact prompt, the reference image (if applicable), and the renderer configuration. Additionally, the VGR rubric used for evaluation must be preserved, as the sample-specific questions are an integral part of the protocol.

It is recommended to repeat runs multiple times to quantify variance and report the uncertainty associated with the score. Comparison between models must be done on the same set of prompts and with the same rendering and evaluation conditions to avoid erroneous conclusions.

Recording the model version, exact prompt, any reference image, renderer settings, and the VGR rubric makes the evaluated setup identifiable. The rubric is part of the protocol, because its sample-specific questions determine which Style properties contribute to the pass rate. Without those details, another run may use different inputs, rendering, or evaluation criteria and produce a result that is not directly comparable.

Repeated runs help characterize variation across generations, while using the same prompts and evaluation conditions makes model comparisons more interpretable. Reporting uncertainty alongside the score communicates that an average over multiple runs is not a guarantee that every generation will receive the same result. These steps support reproducibility, but they do not eliminate sensitivity to the VLM judge or rendering setup.

Checklist to reproduce Style

Steps to ensure the reproducibility of the Style score.

  1. 01Record the version of the AesCode model used.
  2. 02Preserve the generation prompt and reference image.
  3. 03Document the renderer configuration (browser, version, resolution).
  4. 04Save the sample-specific VGR rubric.
  5. 05Repeat the run multiple times to quantify variance.
07

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction