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

Deterministic hint generation (DHT): batch scripts

The scripts in this directory built the hint datasets of the paper. For new images the hintauc library is the easiest entry point (pip install hintauc; hintauc.generate_hints(...) is the same pipeline as a function); the scripts here are the batch tools with sharding and resume.

The pipeline (one colour image → hint maps)

1. Region segmentation — Felzenszwalb (scale=100, sigma=0.5, min_size=100) on the image at its original resolution (the paper's maps were computed from the Danbooru2021 files, not from the 512 × 512 ground-truth copies); SLIC and Quickshift variants (generate_hints.py) and DanbooRegion (danbooregion_hints.py) for the robustness study.
2. Region-id map — a unique colour per region; disconnected parts of one region are split and recoloured.
3. Scribble per region — Zhang–Suen skeleton → 3 × 3 dilation → longest path (FilFinder, branch/skeleton threshold 3 px), clipped to the region; coloured with the region's mean colour.
4. Dot per region — the in-region path pixel with the smallest total Manhattan distance to the other in-region path pixels (dot_method="medoid", the rule of the paper's stored maps).

Output files per image (64 px by default):

<id>.image_region64.png           region-id map
<id>.image_scribble_mask64.png    scribble mask          <id>.image_scribble_col64.png   scribble colours
<id>.image_dot_mask64.png         dot mask               <id>.image_dot_col64.png        dot colours
<id>.image_flatten_img64.png      region-mean colour image

At evaluation time the hints of the largest ⌊α·n⌋ regions are kept (hintauc.HintResult.at_ratio, the DetFill loader). FilFinder breaks medial-axis ties with an unseeded generator, so regenerated scribbles differ from the stored maps (and between two runs) in a few pixels; the stored maps are the reference for every reported number.

Run

pip install -r requirements.txt
python generate_hints.py --segmenter slic --split test \
    --src_root /path/to/src --txt_dir /path/to/split_lists --out_root /path/to/output

• --src_root: <src_root>/segmentation_regions/felzenszwalb/<dir>/<id>.image.png (the colour images).
• --txt_dir: {train,valid,test}.txt, one <dir>/<id>.image.png per line (the paper's lists: ../detfill/configs/illust/).
• Output: <out_root>/<segmenter>/<dir>/<id>.image_{region,scribble_mask,scribble_col}64.png.
• --limit N for a quick check; --shard / --nshards to split the work; --path_method geodesic for the dependency-free longest path (outputs go to <segmenter>_geodesic/; do not mix the two methods in one evaluation).

Felzenszwalb hints for a single image without the batch tool:

import hintauc
hints = hintauc.generate_hints("image.png", size=64)   # Felzenszwalb by default
hints.save("out/image")                                # the file set above

Files