v1.5 — ColorizeDiffusion fine-tuned baselines (200ep and 7ep)

ColorizeDiffusion v1/v2 fine-tuned on the Danbooru2021 training split (illustrations) and on ImageNet (natural images), the baselines of the paper:

• *_200ep: fine-tuned with the common 200-epoch protocol, used for the ColDiff (200ep) rows of Tables II/III and the natural-image tables.
• *_7ep: fine-tuned with the authors' original 7-epoch schedule, used for supplementary Table X and the natural-image tables (7ep).
• Names: coldiff_<v1|v2>_<illust|real>_<dot|scr>_<200ep|7ep>.safetensors; real = ImageNet. coldiff_v1_real_dot_200ep is the last saved state of a run that completed its 200 epochs (saved as latest/).

Each file is split into parts below 2 GiB: cat <name>.part-* > <name>. SHA-256 of the whole files: checkpoints/README.md.
Inference: the official ColorizeDiffusion code with the wrappers described in reproduce/paper_experiments/coldiff_finetune/.

License: these weights are adapted from the official ColorizeDiffusion weights (https://github.com/tellurion-kanata/colorizeDiffusion) and are distributed under CC BY-NC-SA 4.0 (non-commercial, share-alike, attribution), not under the repository's MIT license.

grsi-1.0 — Graphics Replicability Stamp submission

State of the repository submitted for the Graphics Replicability Stamp (GRSI).

• Entry point: bash replicability/run.sh (no arguments) reproduces Figure 9 of the paper; see replicability/README.md for what to expect on different hardware.
• Paper: Hint-AUC: Deterministic Region-based Hint Generation for Line Art Colorization Evaluation, IEEE TVCG 2026, DOI 10.1109/TVCG.2026.3738401.
• Checkpoints and data used by the scripts are the assets of releases v1.0–v1.4 (this release has no binary assets of its own).
• The hintauc library at this state is version 0.3.3 on PyPI.

v1.4 — natural-image models of the paper, earlier-submission and channel-ablation checkpoints

The natural-image models of the paper, the scribble model of the earlier submission and the channel-ablation models.

File	Size	What it is	Notes
detfill_scribble_imagenet_64ch_200ep.pth	485 MB	Natural-image (ImageNet subset) scribble model, 64 base channels	The model behind the supplementary ImageNet tables (configs/scribble_real.yaml). Re-evaluated on 24 test images: PSNR Hint-AUC 18.4 (printed 18.0).
detfill_dot_imagenet_64ch_200ep.pth	485 MB	Natural-image dot model, 64 base channels	The model behind the supplementary ImageNet tables (configs/dot_real.yaml; its conditioning stage has 3 channels, as released). Re-evaluated on 24 test images: PSNR Hint-AUC 15.0 (printed 15.7).
detfill_scribble_illust_earlier_64ch_200ep.pth	485 MB	Scribble model of the earlier submission (64 base channels)	Its archived config samples with 1,000 steps (sample_step: 1000); superseded by the 96-channel v1.0 model retrained for the revision.
detfill_scribble_illust_32ch_200ep.pth	129 MB	Scribble model, 32 base channels, 200 epochs	Supplementary channel-ablation figure ("C = 32"). Use configs/scribble_illust.yaml with model_channels: 32.
detfill_scribble_illust_64ch_200ep.pth	485 MB	Scribble model, 64 base channels, 200 epochs	Channel ablation. model_channels: 64.
earlier_run_configs.tar.gz	2 KB	The archived training configs of these models	argparse-namespace YAML files.

The main tables of the paper use the v1.0 checkpoints. Reproduction example: EXAMPLES_MODE=full bash reproduce/examples/run_examples.sh runs the 32/64/96-channel comparison (E8) and the earlier scribble model (E11) on the example images. SHA-256 sums: SHA256SUMS.txt and checkpoints/README.md.

v1.3 — test-split line art, alternative-segmenter hint maps, example bundle

v1.3 — data for full-scale and example-based reproduction

All files are derived from the Danbooru2021 test split of the paper (3,000 illustrations, ids ending in 016;
test_image_ids.txt inside each archive). The illustrations themselves are not redistributed, except the nine that
appear in the figures of the paper (example bundle); the Danbooru2021 ids are Danbooru post ids, and
reproduce/scripts/fetch_originals.py fetches an image by id and reproduces the dataset's 512×512 copy bit-exactly.
Provided for non-commercial research use to replicate the paper; rights holders can request removal through the
repository contact. SHA-256 of every file: SHA256SUMS.txt.

File	Size	Contents	Used for
test_split_sketch_sketchkeras.tar.gz	441 MB	SketchKeras line art of the 3,000 test images (sketch/sketchkeras/<id>.png, 512×512) — the exact files used for every reported number (sketch type 2)	Tables II/III and all supplementary studies (SketchKeras rows / N=300 studies)
test_split_sketch_pysimp.tar.gz	477 MB	Sketch-simplification line art (sketch/pysimp/<id>.png; sketch type 0)	Tables II/III (mean over the three line-art sources)
test_split_sketch_XDoG.tar.gz	56 MB	XDoG line art (sketch/XDoG/<id>.png; sketch type 1)	same
test_split_hint_maps_64_danbooregion.tar.gz	37 MB	64×64 scribble hint maps and region maps from the DanbooRegion segmenter for the test split	supplementary segmentation-dependency table (evaluation segmenter = DanbooRegion)
test_split_hint_maps_64_slic.tar.gz	38 MB	same from the SLIC segmenter	segmentation-dependency table (SLIC)
test_split_hint_maps_64_felz.tar.gz	44 MB	the Felzenszwalb maps regenerated for that study (the main results use the v1.0 maps)	segmentation-dependency table (Felzenszwalb column of the retrained models)
examples_data.tar.gz	5 MB	the 9 test illustrations that appear in the paper's figures (ids 514016, 4262016, 1625016, 3421016, 4417016, 1019016, 1023016, 4731016, 4942016): originals (512×512), the three line-art versions, the paper's hint maps and the DanbooRegion/SLIC maps	reproduce/examples/run_examples.sh (example-based reproduction of every experiment); other test ids: fetch the originals by id and pass --ids
userstudy_stimuli.tar.gz	606 MB	The images shown in the pairwise user study (Sec. VII): for each of the 192 images and 8 hint ratios the line art, the hint image, and the colorizations of DetFill ("proposed"), PaintsTorch and Diffusart, plus index.csv (imageId → Danbooru id). The original illustrations are not included (fetch them by id from index.csv). Keyed like reproduce/data/userstudy/glmm_trials.csv. The ColorizeDiffusion-v2 stimuli were not preserved (see README.txt inside)	Tables IV/V (inspection of the stimuli behind the trial table)

How to assemble the flat evaluation layout expected by detfill/ (see detfill/README.md, layout A):

mkdir -p DATA_ROOT && cd DATA_ROOT
tar xzf test_split_hint_maps_64.tar.gz            # v1.0: hint_from_regions_64_rev/, region64/
mv test_split_hint_maps_64/hint_from_regions_64_rev . && mv test_split_hint_maps_64/region64 hint_from_regions_256
for t in sketchkeras pysimp XDoG; do tar xzf test_split_sketch_$t.tar.gz --strip-components=1; done   # sketch/<t>/
python reproduce/scripts/fetch_originals.py --id_file test_image_ids.txt --out segmentations/originals   # the 512x512 originals, by id

v1.2 — Diffusart-retrain models, later natural-image checkpoints, ImageNet hint maps

v1.2 — Diffusart-retrain models, later natural-image (ImageNet) models and the ImageNet hint maps

Assets (SHA-256 in SHA256SUMS.txt and checkpoints/README.md)

File	Size	What it is
diffusart_retrain_scribble_dethint_200ep_ema.pth	373 MB	Diffusart (our re-implementation, reproduce/paper_experiments/diffusart_retrain/code/) retrained with the deterministic scribble hints, 200 epochs, EMA weights — the Diffusart-retrain scribble row of the supplementary "additional training comparisons" table (PSNR 19.457 / LPIPS 0.177 reproduced from its per-ratio record).
diffusart_retrain_dot_dethint_200ep_ema.pth	373 MB	Same for dot hints (the dot row of that table).
detfill_scribble_imagenet_96ch_200ep.pth	1.08 GB	DetFill scribble model on the natural-image (ImageNet subset) split, 96 base channels, 200 epochs (detfill/configs/scribble_real.yaml with model_channels: 96). A later training generation: not the model behind the printed ImageNet tables (those are the 64-channel models of v1.4).
detfill_dot_imagenet_96ch_200ep.pth	1.08 GB	Same for dot hints (96 base channels, later generation).
test_split_hint_maps_64_imagenet.tar.gz	40 MB	The stored 64×64 deterministic hint maps (Felzenszwalb regions; scribble/dot colour + mask) and region maps of 2,999 of the 3,000 ImageNet test images (detfill/configs/real/test.txt; n02667379_7324 has no stored map), i.e. the evaluation inputs of the supplementary ImageNet tables; layout as the Danbooru maps of v1.0.

Notes. The Diffusart-retrain models take the 64 × 64 hint maps of v1.0, upsampled to 256 × 256 with nearest-neighbour
interpolation inside infer_det_proposed.py (this is how the paper's record was produced). Our re-evaluation on 24 ImageNet
test images reproduces the printed ImageNet values with the 64-channel models (v1.4) and not with the 96-channel models
released here, which score 2–3 dB higher at every hint ratio. The ImageNet images themselves are not redistributed; the
split lists are in detfill/configs/real/ (relative paths).

Repository changes since v1.1: THIRD_PARTY_NOTICES.md (ColorizeDiffusion-derived inference wrappers stay under
CC BY-NC-SA 4.0; sketchKeras Apache-2.0), the Diffusart re-implementation code, the article DOI
(10.1109/TVCG.2026.3738401) in the README citation.

v1.1 — reproduction package + segmenter-retrained checkpoints

v1.1 — reproduction package and segmenter-retrained checkpoints

Repository (commit 4ceddaa)

• reproduce/: rebuilds the paper's Hint-AUC tables from the released per-ratio metrics (A1_tables_from_released_metrics.py, 322/324 printed cells reproduced), the user-study tables and GLMM statistics (A2_userstudy_glmm.py, all values reproduced), a generic inference + 7-metric Hint-AUC pipeline for the released checkpoints (run_hauc_pipeline.sh, eval_per_ratio.py; HINT_ORDER=label = Table III protocol), training commands (B9_train_detfill.sh), and the verbatim launchers of every experiment of the paper and its revision (paper_experiments/). See reproduce/README.md for the coverage table and the known deviations.
• detfill: dataset_config.hint_order (area = Table II, label = Table III).
• Environment: torchmetrics==1.4.0 pinned (the SSIM implementation behind the published numbers; 1.8.x differs by up to 0.04 per image); perceptual-metric requirements in reproduce/requirements-metrics.txt.

Assets (this release) — DetFill scribble models (96 base channels, 200 epochs) trained on hints from alternative region segmenters; the "DanbooRegion" and "SLIC" training rows of the supplementary segmentation-dependency table. SHA-256 in SHA256SUMS.txt and checkpoints/README.md.

File	Size	Training segmenter
detfill_scribble_illust_danbooregion_200ep.pth	1.08 GB	DanbooRegion
detfill_scribble_illust_slic_200ep.pth	1.08 GB	SLIC

The v1.0 assets (paper checkpoints for scribble/dot and the stored test-split hint maps) remain on the v1.0 release.

v1.0 — paper checkpoints and test-split hint maps

The DetFill checkpoints behind the paper's results, and the stored test-split hint maps.

File	What it is
detfill_scribble_illust_200ep.pth (1.08 GB)	Scribble-hint model, 96 base channels, 200 epochs. Behind every scribble result (Table II, Table III, Fig. 9 scribble row, the supplementary α-grid and seed studies).
detfill_dot_illust_200ep.pth (0.48 GB)	Dot-hint model, 64 base channels, 200 epochs. Behind every dot result (Table II, Table III, Fig. 9 dot row).
test_split_hint_maps_64.tar.gz (60 MB)	The stored 64×64 deterministic hint maps (scribble and dot, colour + mask) and region maps of the 3,000 test images: the exact evaluation inputs of the reported numbers.

SHA-256 sums, configs and usage: checkpoints/README.md.
Place the models as detfill/results/dataset_name/BrownianBridge_{scribble,dot}_illust/checkpoint/latest_model_200.pth, or pass them with --resume_model.