Pixel-art size test: custom model beats GPT Image 2.5
Asked for exactly 16x16 pixel frog sprites, the author's own trained model hits the spec while GPT Image 2.5 and Nano Banana drift on size and quality.
If you ask for pixel art from a frontier image model, it will literally just print at whatever size it feels like. Even with tricks like image edits and references, you can't always get the exact size / aspect ratio you need. I asked for these frogs and included the size in the prompt: "Exactly 16x16 pixels: the whole image is a 16 by 16 pixel grid, one sprite filling it." The model I trained does it perfectly. GPT Image 2.5 and Nano Banana are all over the place. What else is crazy about this comparison is that i kinda think my model might be outperforming these frontier models in quality too. kinda nuts if you think about it.