File size: 2,864 Bytes
8be9ae3 5828b5b 8be9ae3 5828b5b 8be9ae3 5828b5b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 | ---
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
- 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 is a GPU Space that auto-sleeps. The first visit cold-starts the model
> (~30–60s). If the first edit errors, the model is still waking — try again.
## 🔗 Links
- 🎥 **Demo video:** [WATCH ▶](VIDEO_LINK)
- 📣 **Social post:** [SEE POST](SOCIAL_POST_LINK)
- 🧠 **Model (fine-tuned GGUF):** https://huggingface.co/solarkyle/pixellock-gemma-12b-pixelart-gguf
- 📄 **Full technical writeup:** [read it](WRITEUP_LINK)
## 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 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.
|