GraXpert background extraction and denoise in Nebula Studio
=========================================================

GraXpert Development Team
https://github.com/Steffenhir/GraXpert/

Background model: GraXpert 1.0.1, original FP32 ONNX weights, unmodified.
Model licence: Creative Commons Attribution-NonCommercial-ShareAlike 4.0.
Full original notice and contributor credits:
https://irnbrunebula.com/licenses/GraXpert-BGE-Model-LICENSE.html

Model distribution mirror:
https://sourceforge.net/projects/nina-polaris/files/ai-models/bge-ai-models/graxpert-1.0.1/
SHA-256: 4ed3ef68c9f44e821ea4e47572770bb9d4d5df447feae3e758d4c0480fca1f03
Size: 217663463 bytes. No training or weight changes made by Nebula Studio.

Denoise model: GraXpert 3.0.2, original FP32 ONNX weights, unmodified.
Model licence: Creative Commons Attribution-NonCommercial-ShareAlike 4.0.
Full original notice: https://irnbrunebula.com/licenses/GraXpert-Denoise-Model-LICENSE.html
Distribution mirror:
https://sourceforge.net/projects/nina-polaris/files/ai-models/denoise-ai-models/3.0.2/
SHA-256: e61c880e7fb955917e91141660e66801c9de5c4f3d9946dc449e5a7dd7346563
Size: 477703971 bytes. No weight or training changes.

Browser port and processing integration, 15 September 2026:
Adapted from GraXpert's background_extraction.py and denoising.py, commit
c3bc2211f8a4bee864d301987e9a154ad4480969.
GraXpert application source and the adapted integration are GPL-3.0-or-later.
Full licence: https://irnbrunebula.com/licenses/GraXpert-GPL-3.0.txt
Corresponding browser integration source and build instructions:
https://irnbrunebula.com/licenses/studio-graxpert-source.tar.gz
Original Python reference, unmodified:
https://irnbrunebula.com/licenses/GraXpert-upstream-background-extraction.py.txt
https://irnbrunebula.com/licenses/GraXpert-upstream-denoising.py.txt

Changes: browser-only worker inference, WebGPU with CPU/WASM fallback,
verified on-demand model cache, cancellation, zero-MAD guard, reversible
correction strength, low-resolution background surface sampled consistently
for previews and full-resolution exports, mutually exclusive sampled/AI
correction, and explicit rebuild handling for reopened recipes.
Denoise: median/MAD normalisation, 256-pixel context tiles with 128-pixel
centres, copied-edge padding, preserved bright cores and monochrome channels,
zero-MAD/flat-channel protection and bounded native-resolution buffers.
Integration adds retained denoise/starless buffers, signed residual inspection,
reversible blend amounts, original comparison and dependency/rebuild checks.
15 September 2026: denoise changes refresh dependent star separation in an
atomic transaction; cancellation/failure preserves the previous edit, with
one previous native result retained per operation for undo/redo.
This integration is not an official GraXpert application or endorsement.

Runtime: Microsoft ONNX Runtime Web, MIT.
https://irnbrunebula.com/licenses/onnxruntime-LICENSE.txt
https://irnbrunebula.com/licenses/onnxruntime-ThirdPartyNotices.txt

The licence above applies to the GraXpert-derived integration, not the
unrelated server application. Other bundled components retain their notices.