Sking: Image To Minecraft Skin Generator

🌐 Online Demo: EntropyDrop Web Portal
🗂️ Dataset: Hugging Face Sking Dataset
🧠 Trained Model: Hugging Face Sking Model
📝 Technical Report: Technical Architecture
YouTube: YouTube

Sking is an image-to-Minecraft-skin generation model designed to transform character images, portraits, avatars, or concept art into Minecraft-style skin textures. Given a reference image, the model learns to preserve key visual features such as clothing, colors, hairstyle, and character identity while adapting them into the compact pixel-art structure used by Minecraft skins.

SKING_DDJ_v101c — recommended production UV release (2026-09-16)

The portable v101c render-to-UV bundle is available in SKING_DDJ_v101c/. It contains privacy-sanitized inference checkpoints with tensor weights identical to the deployed v101 crown-geometry release, the trained foreground model, pinned offline dependencies, UV mappings, source code, checksums, and a portable launcher.

See the v101c download and inference instructions. Download the complete directory rather than only parser.pt. The older v104 release remains available below for reproducibility.

SKING_DDJ_v104 — production UV release (2026-09-08)

The full v104 render-to-UV bundle is available in SKING_DDJ_v104/. It includes the dense UV parser, trained foreground model, BiRefNet and SigLIP2 dependencies, pinned runtime, UV mappings, and a portable inference launcher. It produces a 64×64 RGBA Minecraft skin from the combined front/back edited render.

See the v104 download and inference instructions. Download the complete directory rather than only parser.pt. Earlier model versions remain available below. The root infer.py is the earlier image-generation entry point; use SKING_DDJ_v104/infer.py for this release.