AI News · Tools

Seven models, one sunflower:
the visual comparison.

DE Auf Deutsch lesen

September 16, 2026 · Reading time approx. 8 minutes

Haiku, Sonnet, Opus, Luna, Terra, Sol and Astra get the same job, blind and isolated: a standalone Blender Python script, officially for Plantwiz, that models and renders the most beautiful, realistic sunflower possible. All seven results were rendered with the same Blender 5.2 installation and are compared here side by side.

Methodology note: Each model got the same prompt in its own fresh session, with no knowledge of the other attempts and no follow-up questions: the most beautiful, realistic sunflower possible, no need to save on vertices, pick matching materials/textures, render with Cycles at 1600×1600. To make the seven images actually comparable, the same camera framing was enforced for all of them (object centered, same viewing angle, same crop) instead of the very different camera positions each model chose on its own. Modeling, color choices, leaf shape, petals, texture and lighting mood are entirely what each model decided for itself.

The brief

All seven models received exactly the same prompt (in German): write a standalone Blender 5.2 script that runs from start to finish with no manual input and produces a complete sunflower (flower head, ray florets, stem, leaves). The brief explicitly stated that the flower should look as beautiful and realistic as possible, that saving on vertices was unnecessary, that matching materials or shader-node textures should be chosen for every part, and that the final result should be rendered with Cycles at 1600×1600 pixels.

All seven, side by side

What makes the difference

ModelColorFlower headLeaves/stem
AstraRich gold-orangeFine spiral structureNaturalistic
OpusPale cream-yellowMost botanically accurate spiralMost detailed (veining)
SolPale creamDense, dark, clear spiralOnly one leaf
TerraSalmon-pink instead of yellowFinely renderedNarrow, jagged
LunaOrange-yellowGreen receptacle stays visibleBamboo-like, segmented
SonnetStrong yellowDark, little structure visibleShadow hides detail
HaikuPale yellowRing with an open centerOne leaf misplaced

A pattern emerges across all seven images: the botanical accuracy of the spiral pattern in the flower head (the most characteristic feature of a real sunflower) and the choice of a rich, warm yellow-orange over a pale or pink-shifted tone separate the most convincing results (Astra, Opus) from the weaker ones (Terra, Haiku). Script length says little on its own: Luna gets by with 108 lines and delivers a coherent, if stylized, result, while Opus writes by far the most elaborate script in the field at 1119 lines and, with it, the most botanically accurate core.

What this means for choosing a model

1

Creative, design-driven tasks reveal real stylistic differences

Unlike pure logic bugs in code, there's no simple right or wrong here. Each model makes its own decisions on color gradients, leaf shape and camera mood, and these decisions differ far more between models than they would on a pure bugfix task.

2

Model size correlates here more with botanical detail

Opus and Astra, the two most elaborate scripts in the field, also deliver the most botanically accurate spiral patterns in the flower head. For tasks where fidelity to a real reference matters, the more capable model tends to pay off here.

3

Without an actual render, this would only have been code reading

All seven scripts read plausibly on their own. Only the actual Blender run with identical camera framing makes visible which model designed the most botanically convincing sunflower, which the most stylized, and which the least coherent.

Honest summary: One brief, seven models, seven visibly different interpretations of the same task. Astra and Opus deliver the most botanically convincing results, Luna and Sol solid, stylized variants, Terra and Sonnet decent but less detailed takes, Haiku the geometrically weakest image in the field. Anyone using AI models for creative 3D tasks shouldn't just check whether a script runs, but actually look at the result.

How reliable is your own AI workflow, really?

I test model and tool choices on real project code, with an actual test run instead of a gut feeling. 30-minute intro call, free of charge.