SKING_DDJ_v101c
Portable open-source package of the v101 production render-to-UV pipeline, published as SKING_DDJ_v101c. It converts a square combined front/back Minecraft character render into a 64×64 RGBA skin with inner and outer layers. The left view is front_left and the right view is back_left, following the Sking edited-render layout. Creating that edited render is a separate upstream stage.
The parser tensor weights, foreground decoder tensor weights, pipeline behavior, and runtime are derived from the deployed v101 crown-geometry release (crown_geometry_20260906, source commit 28f548a5be37868022cfbafaddbe095895e18210). The public checkpoints were repacked to remove optimizer state, private filesystem paths, and development-only manifests. Tensor equality and inference outputs were verified against the deployed release.
Contents
| Path | Purpose |
|---|---|
parser.pt |
Sanitized v101c dense UV parser checkpoint |
foreground.pt |
Sanitized trained Minecraft foreground decoder checkpoint |
pipeline.json |
Inference configuration with development-only path metadata removed |
foreground_base/ |
BiRefNet base model and implementation |
hf_cache/ |
Pinned SigLIP2 model and tokenizer for offline inference |
runtime/ |
Matching inference and training source snapshot, without private test images |
mappings_256x512/ |
UV mapping tensors used by production |
infer.py |
Portable foreground + parser + final inpainting launcher |
release.json |
Checksums, source artifact identities, and sanitization record |
tested_environments.json |
Dependency versions used for verification |
Download the complete directory. parser.pt by itself is insufficient to reproduce the full pipeline.
Download
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="EntropyDrop/Sking",
allow_patterns=["SKING_DDJ_v101c/*"],
local_dir="./Sking",
)
For repeatable deployments, pass the release commit hash as revision instead of following the repository's moving main branch.
Run
Use Linux with an NVIDIA CUDA GPU and Python 3.12. The parser environment needs PyTorch/torchvision, NumPy, Pillow, Transformers, timm, einops, SciPy, OpenCV, SentencePiece, safetensors and accelerate. The foreground environment also uses kornia. Install a CUDA-compatible PyTorch build before the remaining dependencies. The two options below may point to the same interpreter if it contains both dependency sets.
python ./Sking/SKING_DDJ_v101c/infer.py --verify-only
/path/to/parser-env/bin/python ./Sking/SKING_DDJ_v101c/infer.py \
--parser-python /path/to/parser-env/bin/python \
--foreground-python /path/to/foreground-env/bin/python \
--input ./character_edited.png \
--output ./character_result_v101c.png
Input must be an even square image between 256 and 4096 pixels wide, containing the two expected views side by side. The launcher verifies every bundled file, runs the trained foreground model, executes the pinned parser and postprocessing pipeline, validates output provenance, and writes the final 64×64 RGBA PNG. Model inference uses the included Hugging Face cache in offline mode.
Validation and limits
Eight fresh samples were selected across the historical v54, v61, and v66 data groups, excluding the deployment fixtures. Public-package outputs had exact alpha on all eight samples and visible RGB differences of at most 1/255 from the historical v101 results. A separate six-case deployment suite covering eyes, hair, beard, crown, hat, and glasses met the same threshold.
The repacked parser and foreground checkpoints were checked tensor-by-tensor against production. The public package contains no optimizer states, credentials, server paths, user inputs, deployment records, backup directories, or historical regression image galleries.
Semantic errors can still occur, and the model assumes the expected edited-render layout.
License and upstream components
Sking is licensed under AGPL-3.0; see LICENSE. BiRefNet and SigLIP2 retain their own licenses and attribution. See THIRD_PARTY_NOTICES.md and licenses/ for pinned upstream revisions and license texts.