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

Model zoo: released checkpoints and data

Everything needed to run or re-evaluate the paper's models is attached to the GitHub releases of this repository. Each release carries a SHA256SUMS.txt; the tables below repeat the hashes so a downloaded file can be checked with sha256sum -c. Sizes are approximate.

v1.0 — the paper's models (Danbooru2021 illustrations) and their evaluation inputs

Release: https://github.com/MADONOKOUKI/DetFill-with-DHT/releases/tag/v1.0

Where the DetFill launchers expect the two models (detfill/run_inference.sh and reproduce/scripts/run_hauc_pipeline.sh default to this path; replicability/run.sh and reproduce/examples/run_examples.sh download into their own checkpoints/ directories; any location works with --resume_model <file> / CKPT=<file>):

BASE=https://github.com/MADONOKOUKI/DetFill-with-DHT/releases/download/v1.0
mkdir -p detfill/results/dataset_name/BrownianBridge_scribble_illust/checkpoint detfill/results/dataset_name/BrownianBridge_dot_illust/checkpoint
curl -L $BASE/detfill_scribble_illust_200ep.pth -o detfill/results/dataset_name/BrownianBridge_scribble_illust/checkpoint/latest_model_200.pth
curl -L $BASE/detfill_dot_illust_200ep.pth      -o detfill/results/dataset_name/BrownianBridge_dot_illust/checkpoint/latest_model_200.pth

Notes - Both checkpoints hold the EMA weights saved by the training loop at epoch 200; DetFill is a pixel-space model, so no VQGAN or latent-diffusion weights are needed. - Configuration: detfill/configs/scribble_illust.yaml (96 channels) for the scribble model, detfill/configs/dot_illust.yaml (64 channels) for the dot model. Inference is seeded (--seed 1234, the default of detfill/main.py). - The hint maps in test_split_hint_maps_64.tar.gz are the reference inputs. They were generated from the Danbooru2021 files at their original resolution (not from the 512 × 512 ground-truth copies); regenerating them from those files with the hintauc library reproduces the region count and almost all scribble pixels (mean mask IoU 0.992 over the 9 example images), but FilFinder's unseeded tie-breaking moves a few pixels per run (reproduce/examples/README.md, E9), so use the stored maps whenever you want to compare with the printed numbers. - Metric backbones (LPIPS, OpenCLIP, DINOv2, DreamSim) are downloaded by their pip packages on first use and are not part of the releases. - The ImageNet hint maps of release v1.2 cover 2,999 of the 3,000 ids in detfill/configs/real/test.txt (n02667379_7324 is missing); reproduce/scripts/check_data_root.py reports such gaps before an inference run starts, and the loader reads the whole list, so remove missing ids from the list (or add the files) first.

v1.1 — scribble models trained with other region segmenters

Release: https://github.com/MADONOKOUKI/DetFill-with-DHT/releases/tag/v1.1 — the "DanbooRegion" and "SLIC" training rows of the supplementary segmentation-dependency table (same protocol and training split as the v1.0 scribble model, 96 base channels, 200 epochs; only the segmenter that produced the training hints differs).

Use them with detfill/configs/scribble_illust.yaml and --resume_model <file>; the evaluation data root must contain hint maps from the matching segmenter (release v1.3, test_split_hint_maps_64_{danbooregion,slic}.tar.gz).

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

Release: https://github.com/MADONOKOUKI/DetFill-with-DHT/releases/tag/v1.2

The Diffusart-retrain models take the 64 × 64 deterministic hint maps of v1.0, upsampled to 256 × 256 with nearest-neighbour interpolation inside the inference script — this is how the paper's Diffusart-retrain record was produced. Inference: reproduce/paper_experiments/diffusart_retrain/code/infer_det_proposed.py (PyTorch 2.5 and diffusers 0.32; see the README in that directory).

Natural-image models: two training generations exist. The 64-channel models (release v1.4) reproduce the printed ImageNet values (24 test images, SketchKeras line art: PSNR Hint-AUC 15.0 dot / 18.4 scribble against 15.7 / 18.0 printed), whereas the 96-channel models released here score 2–3 dB higher at every hint ratio; the printed supplementary ImageNet tables therefore come from the 64-channel models, and the 96-channel ones are a later, stronger generation that is not in the paper. The *_real.yaml configs use the split-based layout (dataset_path plus the relative lists in detfill/configs/real/, run main.py from detfill/); the ImageNet images themselves are not redistributed.

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

Release: https://github.com/MADONOKOUKI/DetFill-with-DHT/releases/tag/v1.3 (all files derived from the Danbooru2021 test split; the originals are not included — they are available from https://gwern.net/danbooru2021 by id; non-commercial research use).

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

Release: https://github.com/MADONOKOUKI/DetFill-with-DHT/releases/tag/v1.4. The two natural-image models here are the ones behind the supplementary ImageNet tables; the other files are released for completeness and for the supplementary channel ablation and are not behind the main tables (those are v1.0).

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

Release: https://github.com/MADONOKOUKI/DetFill-with-DHT/releases/tag/v1.5. The ColorizeDiffusion v1/v2 models of the paper's comparisons, fine-tuned from the official weights: *_200ep with the common 200-epoch protocol (the ColDiff rows of Tables II/III and of the ImageNet tables), *_7ep with the authors' original 7-epoch schedule (supplementary Table X and the ImageNet tables). Each file is split into parts below 2 GiB; restore it with cat <file>.part-* > <file> and check the SHA-256 below. Run them with the official ColorizeDiffusion code and the wrappers in reproduce/paper_experiments/coldiff_finetune/. License: CC BY-NC-SA 4.0 (adapted from the official ColorizeDiffusion weights), not the repository's MIT license.

Which release do I need?