Mirror of the documentation in the DetFill-with-DHT repository (source file THIRD_PARTY_NOTICES.md); links to files and release downloads point at GitHub, and the hintauc library is available from PyPI.

Third-party code and licenses

The repository is released under the MIT License (LICENSE), with the exceptions and attributions below.

Location	Upstream	License	Notes
`detfill/`	BBDM — Brownian Bridge Diffusion Models (Li et al., 2023)	MIT (`detfill/LICENSE`)	DetFill is a fork adapted to sketch + deterministic-hint conditioning.
`reproduce/paper_experiments/coldiff_finetune/coldiff_v_inference_official_fixed.py`	ColorizeDiffusion (Yan et al.)	CC BY-NC-SA 4.0 (`reproduce/paper_experiments/coldiff_finetune/LICENSE`)	Modified copies of the official inference script (deterministic-hint input). They are distributed under the upstream license, not MIT, and only run inside the upstream code base with its weights.
Release v1.5 assets `coldiff_*.safetensors`	ColorizeDiffusion (Yan et al.)	CC BY-NC-SA 4.0	Our fine-tuned versions of the official ColorizeDiffusion v1/v2 weights. Distributed under the upstream license (non-commercial, share-alike, attribution), not MIT.
`reproduce/scripts/sketch_tools/sketchkeras_*.py`	sketchKeras (lllyasviel)	Apache-2.0 (`sketchkeras_LICENSE`)	Our batch wrapper around the upstream helper; the `mod.h5` weights are downloaded from the upstream release.
`reproduce/scripts/sketch_tools/aki_edit_d.py`, `pysketchsimplify_*.py`	Sketch Simplification (Simo-Serra et al.)	MIT (code); pretrained weights for non-commercial research use only	Model definition and batch wrapper used to produce the "sketch simplification" line art; obtain the pretrained model from the upstream repository under its terms.
`reproduce/scripts/sketch_tools/XDoG_*.py`	XDoG (Winnemöller et al., 2012)	—	Our own NumPy/OpenCV implementation used for the XDoG line art.
`reproduce/paper_experiments/segmenter_dependency/r2-2_scripts/D_danbooregion_seg.py`	DanbooRegion (Zhang et al., 2020)	upstream	Driver script only; it imports the upstream code and weights, which are not included.
`reproduce/paper_experiments/diffusart_retrain/`	our re-implementation of Diffusart (Carrillo et al., CVPRW 2023)	MIT (this repository)	Training/inference code written in our lab for the paper's Diffusart baseline and its deterministic-hint retraining; no upstream code is included.
`hintauc/`, `evaluation/`, `hint_generation/`	—	MIT	Uses FilFinder, scikit-image, LPIPS, OpenCLIP, DINOv2 (transformers), DreamSim and torchmetrics as pip dependencies; their weights are downloaded by the respective packages under their own licenses.

Runtime dependencies and pretrained weights

Dependency	License	Used for
FilFinder (BSD-3), astropy (BSD-3), scikit-image (BSD-3), OpenCV (Apache-2.0), NumPy / SciPy / Pillow (BSD / HPND)	free	hint generation, pixel metrics
PyTorch, torchvision (BSD-3); torchmetrics (Apache-2.0); torch-fidelity (Apache-2.0)	free	DetFill, SSIM / MS-SSIM, FID / KID
LPIPS (BSD-2; AlexNet weights ship with the package)	free	LPIPS, LPIPS-VGG
OpenCLIP (MIT) with the LAION-2B ViT-B/32 weights (MIT)	free	OpenCLIP metric
DINOv2 through `transformers` (Apache-2.0; weights Apache-2.0)	free	DINO metric
DreamSim (MIT; weights MIT)	free	DreamSim metric
DISTS (`dists-pytorch`, MIT)	free	DISTS metric
taming-transformers (MIT), vendored under `detfill/model/VQGAN/taming` by BBDM	free	not used by the released configs
Contributor Covenant 2.1 (CC BY 4.0)	—	`CODE_OF_CONDUCT.md`

All of them allow free non-commercial and academic use; weights are downloaded by the respective packages on first use.

Data

The illustrations of the paper come from Danbooru2021 (Danbooru posts; the Danbooru2021 ids are the post ids). The images themselves are not redistributed, with one exception: the illustrations that appear in the figures of the paper. Everything else refers to the original distribution by id, and reproduce/scripts/fetch_originals.py fetches an image by id from Danbooru and reproduces the dataset's 512 × 512 copy bit-exactly.

Released for non-commercial research use only, so that the paper can be replicated:

derived data of the 3,000 test images, without the images: the 64 × 64 hint maps and region maps (v1.0, v1.3), the line art (v1.3) and the per-ratio metric files (reproduce/expected/);
the nine paper-figure illustrations of the example bundle examples_data.tar.gz (v1.3; ids 514016, 4262016, 1625016, 3421016, 4417016, 1019016, 1023016, 4731016, 4942016) with their line art and hint maps, the same two Fig. 9 illustrations in replicability/data/ (see replicability/data/NOTICE.txt), the demonstration images in assets/readme/ (ids 4262016 and 514016), the paper's archived outputs and the authors' example grids in reproduce/examples/expected/ (these nine images only);
the user-study stimuli (v1.3): line art, hint images and the colorizations of three methods, indexed by index.csv; the original illustrations of the study are not included (fetch them by id from index.csv).

Rights holders can request the removal of an image through the repository's issue tracker or the contact in replicability/GRSI_SUBMISSION.txt. The natural-image experiments use an ImageNet subset; only its 64 × 64 hint maps are released (v1.2).