Haiku 5.5 Meets the Portrait Benchmark

Haiku 5.5 is out, so I put it through the portrait benchmark from Teaching Coding Agents to Paint. A coding agent paints a woman's portrait in code, with one brush: splat(), a soft-edged ellipse on a 384 × 512 canvas. Greys only, at most 300 strokes.

The first post covers many models and modes. Haiku 5.5 got two of them, with the original prompts: one-shot, and canonical.

One shot

The agent writes the whole painting in a single pass and never sees it. Haiku 4.5 did this in the first post, so here are both Haiku models, side by side.

Neither used much of its budget. Haiku 4.5 painted a soft, cartoonish face. Haiku 5.5 planned a three-quarter portrait lit from the left, with one bright earring to pull the eye to the lit cheek, and the result reads as a woman, though a dim one.

Canonical

This is the hardest exercise. The agent paints in ten stages, from 5 strokes up to 300, and renders its work after each one. It also has to pick a canon of proportions before the first stroke, and place every stroke of the head from that construction instead of by eye.

Haiku 5.5 chose a Loomis-style canon, lit the head from the upper left, turned it a little, and gave the woman long wavy hair.

It stumbled. The head was too narrow at stage 1, and a bug in stage 2 collapsed the neck into a ring. The hair was worse: after stages 7 and 8 the agent saw that it looked like dots floating above the skull, but the rules allow one render per stage, so it could only fix it in stage 9, and not completely.

The final head is solid and the features sit where the construction puts them. The eyes win the eye, as intended. The hair and shoulders stay weak, which the agent admits in its own notes.

Verdict

One run per exercise proves little. Still, blind, Haiku 5.5 looks better than Haiku 4.5 to my eye, and in the long guided exercise it spotted its mistakes more easily than it fixed them. The first post has the other models for comparison.