You can deploy this model and start earning money today!

solarkyle/
pixellock-gemma-12b-pixelart-gguf

Name: pixellock-gemma-12b-pixelart-gguf
Availability: InStock

grammar-constrained-decoding

Views

No views yet

PixelLock — Gemma-4-12B Pixel-Art Retexturer (GGUF, q4_k_m)

A QLoRA fine-tune of google/gemma-4-12b-it for footprint-locked pixel-art retexturing.

It is a text LLM, not an image generator. A sprite is serialized to a palette-indexed text grid (a PALETTE block plus a GRID of space-separated cells). The model rewrites that grid in a new style, and a per-sprite GBNF decoding grammar (llama.cpp) forces every transparent cell to stay transparent — so the exact silhouette and transparency are preserved by construction, not by luck.

What it does

Retheme / restyle palette-indexed pixel-art sprites (best at 16-64px) to any vibe (lava, ice, gold, cosmic, ...) - pixel-perfect.
2x upscale with added shading (each input pixel becomes a 2x2 block).

How to use

Run with llama.cpp and pass a footprint-derived GBNF grammar per sprite. The sprite is sent as a PALETTE + GRID wire block; the grammar pins every transparent cell. See the PixelLock app for the wire format and grammar builder.

Training

Supervised fine-tune (QLoRA, rank 64, alpha 64, lr 1e-4, completion-only loss) on a curated corpus of palette-indexed pixel-art sprites (<=64px), with step-checkpointing + best-eval selection and a held-out eval set. Final eval loss ~0.38.

Files

pixellock-gemma-12b-q4_k_m.gguf - q4_k_m quant for llama.cpp / LM Studio / Ollama.

Built for the Build Small hackathon.

solarkyle/pixellock-gemma-12b-pixelart-gguf