OpenAI

Tests multi-subject scene composition and a spatial relationship that’s easy to get wrong (boy outside, cat inside, looking through glass). The whole point is warmth and charm, not precision, so it rewards clean appealing rendering over technical accuracy.

Tests restraint and recognizability at icon scale: Can the model say a lot with few shapes and resist over-decorating? This is where compactness genuinely matters, since “minimalist / flat / clean lines” is the explicit brief.

It’s hard because it forces a model to combine the rigid mechanical geometry of a bicycle with the organic anatomy of a pelican and then compose them into a coherent scene where the bird is actually pedaling. Even leading models frequently get the frame structure wrong or misplace the pelican’s limbs, which makes it a surprisingly revealing test of spatial reasoning.