The prompt is deceptively simple: Generate an SVG of a frog with a Habsburg jaw. To a casual observer, this looks like a typical piece of AI prompt engineering designed for a laugh. However, for developers and AI researchers, this specific request serves as a surgical tool to probe the boundary between strict instruction following and autonomous narrative expansion. By asking a model to render a highly specific, historically charged physical trait in a structured code format, researchers can observe exactly where a model stops being a tool and starts acting as an editor.

The Metrics of Visual Code Generation

The benchmark focused on three primary dimensions: latency, output size, and the nature of the annotations embedded within the SVG code. Each model underwent three iterations of the same prompt to ensure consistency. The results revealed a staggering variance in performance. Response times ranged from a lean 5.4 seconds to a sluggish 257.1 seconds. This gap suggests that different models are employing vastly different internal strategies to process the request, with some treating it as a simple geometry task and others treating it as a creative writing exercise.

File sizes mirrored this disparity, spanning from a minimal 536B to a dense 14,329B. The data indicates a direct correlation between the size of the output and the depth of the accompanying annotations. Models that produced larger files did not necessarily create more complex visual shapes, but they did inject significantly more metadata into the code. These annotations generally fell into three distinct tiers of complexity. The first tier consisted of basic structural labeling, where the model simply identified body parts such as Eyes or Legs. The second tier moved into editorial interpretation, using descriptive phrases like massive protruding mandible to explain the visual implementation of the Habsburg jaw.

The third and most complex tier involved the integration of external knowledge. In these instances, models added elements that were never requested in the prompt. Some models decided the frog needed a Royal Imperial Collar or a Golden Fleece Medal, effectively transforming a simple animal drawing into a historical portrait. This suggests that the prompt triggered a wide-reaching retrieval process, pulling from historical datasets regarding the House of Habsburg to fill in the gaps of the visual representation.

The Correlation Between Latency and Over-Editorializing

The most revealing aspect of the benchmark is not the visual output, but the cognitive leap models take when they over-editorialize. There is a clear divide in how models interpret the term Habsburg jaw. Lower-performing or faster models treated the term as a literal description of a physical deformity, often labeling the trait as a Habsburg jaw - extreme underbite. These models stayed within the guardrails of the prompt, prioritizing speed and structural accuracy over depth.

In contrast, high-latency models exhibited a phenomenon that can be described as narrative expansion. Instead of simple descriptions, these models utilized precise medical terminology, labeling the jaw as THE PROMINENT HABSBURG JAW (PROGNATHISM). The use of the term prognathism indicates that the model is not just recalling a visual pattern but is accessing a specialized medical lexicon to justify its design choices. This shift from description to definition marks a transition in the model's internal reasoning process.

This expansion often spiraled into full-scale historical hallucinations. Some models began adding annotations that described the frog's personality and social standing, including labels like Royal Velvet Cushion, Haughty Expression, and even Inbred Royal Disdain. These additions are not visual elements but textual markers that reveal the model's internal associations. The model is essentially telling the developer that it perceives the Habsburg jaw not just as a shape, but as a symbol of aristocratic decay and genetic isolation.

The data shows that this narrative depth comes at a significant computational cost. Models that spent over 100 seconds generating the SVG were the ones most likely to produce these complex, historically informed annotations. The low-latency models, typically responding in under 10 seconds, almost entirely avoided this behavior, sticking to basic structural labels or minor emotional cues like a Smirk. This creates a clear trade-off: the more a model attempts to synthesize external context and provide a sophisticated interpretation, the slower the output becomes.

For developers, this reveals a critical tension in LLM deployment. When an AI is used to generate structured assets like SVGs, the goal is usually precision and adherence to specifications. However, the very capabilities that make a model feel intelligent—its ability to connect a jawline to the history of the Holy Roman Empire—become liabilities in a production pipeline. This over-editorializing can lead to inconsistent outputs where one asset is a simple frog and the next is a fully realized royal caricature, all from the same prompt.

Choosing the right model for SVG generation therefore requires a balance between efficiency and creativity. If the objective is the rapid production of clean, predictable assets, a low-latency model with minimal annotation tendencies is superior. If the goal is character design or conceptual art where the AI is expected to bring its own creative intuition to the table, the high-latency, narrative-driven models provide a level of depth that simple instruction-following cannot match.

Ultimately, the Habsburg jaw benchmark proves that annotations in code are a window into the model's latent space. By analyzing the gap between what was asked and what was annotated, developers can quantify a model's tendency toward hallucinatory expansion and decide whether that creativity is an asset or a bug in their specific use case.