I trained a sprite model with agents. The data was the bottleneck.

I just published pixel-llm, a small autoregressive transformer that generates 32x32 pixel art sprites of reef sea creatures. About 2.9M parameters, a 64-colour palette, runs on consumer hardware. Built end to end through agent sessions, with me steering rather than typing. The output is sub-par. I am sharing it anyway, because the way it failed taught me something I did not expect. The setup was narrow on purpose.