Can you ask AI to make a penguin?
Everything you need to practise: the penguin references, the final image, the scoring script and the rules. Version 0.2 (21 Oct) fixes the colour score and the camera notes.
What is in the zip
| File | What it is |
|---|---|
penguin/front.png, side.png, back.png, top.png | The four reference views of the practice animal. |
penguin/final.png | The finished look, in a scene. The experts judge your texture against it. The script does not use it. |
penguin/*-mask.png | Silhouettes (white = animal) that the scorer compares your model with. The reference views are cut-outs on a transparent background. |
penguin/subject.json | Real size in metres: height 0.70, length 0.30, width 0.25. |
score_glb.py | The scoring script. It needs Python, numpy and Pillow. |
README.md | Input and output formats, the camera directions (the nose points towards -y), the scoring rules, and what the score cannot see. |
Score your own model
python score_glb.py out/untextured.glb penguin/ --out score.json python score_glb.py out/textured.glb penguin/ --out score.json --textured
Shape: the overlap of your model's silhouette with the reference silhouette in the front, side and top views, taking the best two of the three. The scorer also lists gross errors (a missing head or foot, a nose at the wrong end). Colour: your textured model compared with the reference views, background removed, again the best two of three. To pass the judging you need 0.80 on shape with no gross errors, and 0.80 on colour, on each of the three hidden animals.
A model of the wrong thing still scores around 0.5 on shape. The script only sees outlines, so the experts also check that your model is complete and recognisable. The reference pictures are AI-made and not perfectly consistent (for example, the penguin's back-view feet face the wrong way). We will not repair them; handling that is part of your skill.
Rules in brief
- You write an agent skill (a
SKILL.mdplus scripts). The skill is what is judged, not a model you made by hand. - Allowed models, at these versions or lower: Opus 5.5, Sonnet 5.5, Sol (gpt-6.1-sol), Luna (gpt-6-luna). These are the only AI models allowed. AI image generation is allowed, including to fix, repair or create images. Not allowed: AI 3D generation (anything that produces a 3D shape, mesh or scene from text or images) and using an existing 3D model of the animal as reference.
- Staff run your skill unattended on three hidden animals. They are not penguins and may look very different: another body plan, thin or soft parts, no legs.
- Each judged run is limited to 2.5 million tokens and 2 hours. Your own testing is not limited.
- No hand edits to any output.
- Test your skill on other animals before you submit, for example a turtle, a frog, a giraffe, a bat or a crab. A skill that only works for penguins will fail.
- Judging is the script score, then expert review. The reward goes to the first team whose submission passes both, by submission time.
- Up to 2 submissions per team.
Dates
- Friday 23 October 2026: a team of several students registers by email with member names and student IDs. A student working alone does not register.
- Tuesday 10 November 2026, 11:59 pm Eastern: submission deadline.
Questions and registration: randy.lin@algomau.ca
The practice penguin
