Home / Benchmark / illustration

Three models got this style wrong. All seven got the colours right.

Illustration is the hardest shape target in our set: overlapping figures, irregular crops, rounded frames sitting on rounded cards. We gave seven models the same DESIGN.md, the same copy and the same output contract, and ran each one three times. Colour fidelity was 100% for every model. Shape fidelity was not.

12.7%Qwen3.8-Max, lowest shape fidelity of any model on any of the six benchmark styles
3 / 7models below 75% shape fidelity
100%palette fidelity, every model
73pplargest within-model range, Doubao Seed 2.1 Pro — three runs, one of them 71 points below the others

Every model, same spec, three runs

Shape fidelity is the share of border-radius and box-shadow values in the output that match the spec. Range is the spread between the best and worst of the three runs for that model. Cells below 75% are shaded. Seven open-weight models, each on its own vendor endpoint, no aggregator, temperature 0.7, thinking disabled, three independent samples.

ModelShape (median)RangePaletteType
Qwen3.8-Max 12.7% 43pp 100.0% 100.0%
Doubao Seed 2.1 Turbo 24.3% 28pp 100.0% 100.0%
GLM-5.3 63.6% 42pp 100.0% 100.0%
Qwen3.8-Flash 81.2% 38pp 100.0% 100.0%
Doubao Seed 2.1 Pro 100.0% 73pp 100.0% 100.0%
DeepSeek V4.1 Flash 100.0% 60pp 100.0% 100.0%
Kimi K3.1 Pro 100.0% 0pp 100.0% 100.0%

What breaks

Three places where this style leaves the model something to decide. Each one is a gap in the specification, not a reading error by the model.

The finding

Illustration breaks on geometry, never on colour. Three models land below 75% shape fidelity while hitting 100% on the palette — the spec is not misread, it is under-specified where the shapes overlap.

If you build in this style, the checks that matter are on radii and shadow layers, not on the palette.

Coverage and method. This page covers one of six benchmark styles, the only six for which all seven models were run. The other 22 styles in the library were measured on a single model (Doubao Seed 2.1 Turbo, the one run across all 28 styles), so their numbers describe that model’s variance and are not cross-model comparisons. Every figure here comes from a machine check of the generated HTML against the spec’s token table — no human rating, no screenshots. Raw requests and responses are archived; how it is measured.

Benchmark index · See this style rendered · All 28 specs