⚡ NOT an image generatoryour PNG → a text grid (PALETTE + GRID) → a fine-tuned text LLM rewrites it under a GBNF grammar that locks the exact shape → pixel-perfect PNG. It's language modeling, not diffusion.
⌖ see what the model actually wrote — it's text, not pixels
Run an edit, then peek here: the LLM literally writes a PALETTE block and a GRID of cells (K R R K …). The GBNF grammar forces every transparent cell to stay transparent — that's why the shape can't break.