---
title: PixelLock
emoji: 🎨
colorFrom: purple
colorTo: indigo
sdk: docker
app_port: 7860
pinned: true
license: gemma
short_description: Footprint-locked AI pixel-art retexturing
models:
  - solarkyle/pixellock-gemma-12b-pixelart-gguf
tags:
  - gradio
  - build-small-hackathon
  - track:wood
  - sponsor:modal
  - achievement:offbrand
  - achievement:welltuned
  - achievement:llama
  - achievement:sharing
  - achievement:fieldnotes
  - achievement:offgrid
  - best-demo
  - bonus-quest-champion
  - pixel-art
---

# 🎨 PixelLock — AI Pixel-Art Retexturing

Upload a sprite, pick a vibe (or write your own), hit **Run**. PixelLock restyles
the colors and shading while keeping the **exact silhouette** — pixel-perfect,
*guaranteed by construction*.

> ⚠️ **This Space runs on CPU** (the hackathon org has no dedicated-GPU credits), so
> live edits won't run *here*. **▶ Try it live on GPU →
> https://huggingface.co/spaces/solarkyle/pixellock** (cold-starts ~60s) — or watch
> the 45-second demo video linked below.

## 🔗 Links

- 🎥 **Demo video:** [WATCH ▶](https://youtu.be/Ju-Mpddfn3c)
- 📣 **Social post:** [SEE POST](https://x.com/NeuronsAndBytes/status/2066577400834506843)
- 🧠 **Model (fine-tuned GGUF):** https://huggingface.co/solarkyle/pixellock-gemma-12b-pixelart-gguf
- 📄 **Full technical writeup:** [read it](https://huggingface.co/solarkyle/pixellock-gemma-12b-pixelart-gguf/blob/main/TECHNICAL_WRITEUP.md)

## The idea: the *wrong* model on purpose

PixelLock does **not** use an image generator. It uses a **fine-tuned text LLM**
(Gemma-4-12B). The sprite is serialized to a tiny text wire — a `PALETTE` block
plus a `GRID` of space-separated cells — and the model **rewrites that text** in a
new style. The trick is a per-sprite **GBNF decoding grammar** (llama.cpp): every
transparent cell is pinned to a fixed literal, so the model is *physically
incapable* of moving a pixel. The footprint and transparency are preserved **by
construction**, not by luck. It's language modeling, not diffusion — and that's
exactly why the shape can never break.

In a benchmark, *no* model could emit a correct 32×32 grid unconstrained
(~0/150 attempts); the grammar takes validity from **0% → 100%** at every size
(16×16 → 128×128, incl. 2× upscale), footprint-perfect every time.

## The tech

- **Fine-tuned model** — QLoRA fine-tune of Gemma-4-12B (**trained on Modal**,
  A100-80GB) on a curated corpus of palette-indexed pixel-art sprites (≤64px);
  completion-only loss, step-checkpointing, final eval loss **0.378**. Exported
  to GGUF (q4_k_m).
- **Grammar-constrained decoding** — a per-sprite GBNF grammar compiled from the
  input footprint; the silhouette is locked in the decoder, not post-filtered.
- **Custom Gradio UI** — a heavily themed `gr.Blocks` interface (Off-Brand) that
  builds the grammar per upload, footprint-checks the output, and shows you the
  **raw text the model wrote**.
- **Serving** — Docker GPU Space running llama.cpp `llama-server` (CUDA) with the
  per-request grammar; auto-sleeps when idle.

Built for the **Build Small** hackathon · *Thousand Token Wood* track.