Kairo 3D Bake-off
Dragon STEM · 3D pipeline · Kairo pilot

Kairo 3D Bake-off

Ten OpenRouter image models drew Kairo's turnaround from his character sheet, in 23 variants. Every set of drawings went through the same local Hunyuan3D-2mv run on an RTX 4090, was painted from its own drawings, and was scored by six AI judges on likeness, turnaround consistency and 3D production quality. Pick any entry to spin its 3D model.

Kairo's character sheet: turnaround, expressions, poses and locked identifiers
Source: Kairo character sheet

The five drawings it made (front, left, back, right, three-quarter), then its 3D model rendered front, left, back and three-quarter
What the bake-off showed

Findings

One model drew a true turnaround

GPT Image 2.5 Sunburst was the only model whose drawings were clean orthographic views on white, with Kairo's brass cuff on the left foreleg in all five views. It took 1st, 2nd and 4th place.

Edge-on wings fix the side views

Spread wings drawn as a raised sheet in profile contradict the front view. Redrawing the side views with the wings edge-on lifted Sunburst from 76.9 to 87.3 and GPT Image 2 from 49.2 to 67.8. It only helps when a model obeys on both sides; several added a second cuff.

The 2×2 grid is the cheap, on-model surprise

Asking Sunburst for all four views in one image gave the most compact, book-like Kairo for $0.10 (4th, 71.3), but at half the resolution and with no three-quarter view to paint from. Worth a follow-up test.

Best non-OpenAI option

Gemini 3.1 Flash Image (3rd, 71.9) gave the sharpest drawings, the cleanest single cuff and the flattest wedge tail, with muted colours and a longer neck.

Models to avoid for turnarounds

FLUX.2 Max, Grok Imagine and Gemini 3 Pro (one view at a time) drew two-legged or T-pose front views, extra dragons or cuffs on several legs, and the 3D generator faithfully built those extra limbs. The prompt matters too: the first prompt scored 23.7, the current one 71.9, both with Gemini 3.1 Flash.

Numbers need eyes

Silhouette overlap (IoU) alone misleads: a six-limbed Grok model scored 0.69. The judges' close-up checks of horns, eye dashes, claws and the cuff decided the ranking. They also caught two pipeline bugs, which were fixed and every mesh regenerated.

For the other ten dragons

Recommendation

Use GPT Image 2.5 Sunburst, one view at a time, with edge-on side views: about $0.43 per dragon, roughly $5 for all ten heroes and villains.

  • Add each sheet's proportions to the prompt ("tail no longer than the body, short torso"): Sunburst draws Kairo longer than the book, with an arrowhead tail tip instead of the flat wedge.
  • Check every drawing automatically before meshing: one-sided identifiers (Kairo's cuff, Zephyra's ribbon, Terra's vials, Lumen's lens) appear once and on the correct side, and there are no two-legged views, extra characters or clipped wingtips.
  • Fallback when a dragon fails twice: Gemini 3.1 Flash Image, one view at a time.
Same recipe for every candidate

How each 3D model was made

  1. DrawThe image model draws the front view from the sheet, then the left, back, right and three-quarter views using that front view as an anchor.
  2. Cut outBackground removed; all four orthographic views scaled to the same height so they describe one object.
  3. BuildHunyuan3D-2mv (local, RTX 4090) builds a mesh from the four views, twice with different seeds.
  4. Score & paintThe mesh's outline is compared with each drawing; the better seed is painted from the drawings it can see.
  5. JudgeSix judges (two per lens) score likeness, turnaround consistency and 3D production quality; a synthesis ranks all 23.