The launchers that were actually run
The shell and Python files that produced the paper's and the revision's numbers, copied from our lab tree. They are example code: cluster paths and host names are replaced by the placeholders
/scratch/USER/..., /home/USER/... and HOST_A–HOST_D, conda environment names (BBDM, py310) and GPU ids are hard-coded, several were
written for a specific host queue, and they are not maintained as runnable entry points. The runnable, path-free
equivalents are ../scripts/run_hauc_pipeline.sh, ../scripts/eval_per_ratio.py, ../scripts/A1_* / A2_* and
../examples/. Directory names with "R1-1", "R2-2" and so on refer to the reviewer comments of the first TVCG revision
(reviewer number – comment number) that the experiment answered.
| Directory | Paper item | Contents |
|---|---|---|
tableII_inference/ |
Table II (DetFill rows) | run_inference_mr.sh: the loop over 8 hint ratios × 3 line-art sources (main.py --sample_to_eval) for the dot and scribble models. Released as detfill/run_inference.sh. |
table3_label_order/ |
Table III (DetFill rows) | launch_tableIII.sh: the same loop with the fixed ascending-label region order (hint_order: label). |
alpha_grid/ |
Table II scribble row; supplementary hint-ratio-grid (α-grid) table and dense-curve figure | B_run.sh / B_run_parallel.sh → B_eval_dense_curve.py (per-image metrics over the 57-ratio sweep) → B_build_summary.py (per-ratio summary) → B_auc_grid_sensitivity_7m.py and B_vary_N_auc.py (Hint-AUC on alternative grids) → B_plot_dense_curve.py. A-5_alpha_grids.py: user-study win-rate AUC under alternative grids (review response). Results: ../expected/alpha_grid/. |
seed_sensitivity/ |
Supplementary "size-ordered vs. random selection" table and figure | gen_seeded_random_hints.py (seeded random region order baked into 64 × 64 maps), build_wrappers.py (symlinked data roots per seed and ratio), run_seedexp_{detfill,diffusart,coldiff}.sh, watcher_launch.sh (queueing), detfill_cfgs/ (the generated configs), eval_n300/ (eval_sweep.sh + the seven-metric evaluator restricted to the 300-image SketchKeras subset; the size-ordered counterpart used the same scripts on the deterministic hints). Results: ../expected/seed_sensitivity/. |
segmenter_dependency/ |
Supplementary segmentation-dependency table and figures | 10_run_ratio.sh / 20_launch.sh / 30_push_and_eval.sh (per-ratio inference across a GPU pool, copy to the file server, evaluation), configs/ (training configs of the 96-channel model and the eval_on_{felz,danbooregion,slic}.yaml evaluation configs), r2-2_scripts/ (DanbooRegion/SLIC segmentation and hint generation: D_danbooregion_seg.py, D_danboo_hints.py, D_retrain_gen_hints.py; retraining: D_train_launch.sh, D_stage_train.sh; cross-evaluation matrix: D_eval_matrix.sh; figure montage: D_build_allratio_figs.py, D_montage.py; C-1_segmenters.py, C-2_perturbation.py = hint-containment and perturbation analyses of the review response), R2-2_README_original.md. Checkpoints: release v1.1; test hint maps: release v1.3. Results: ../expected/segmenter_dependency/, ../expected/segmenter_sensitivity/. |
segmenter_sensitivity/ |
Review response | C-1_segmenters.py, C-2_perturbation.py (same as above, original location). |
userstudy_rank_stability/ |
Review response | A-1_bootstrap.py (per-participant bootstrap of win rates and rank flips), A-2_binomial.py, A-3_bradley_terry.py; they read the raw per-participant CSVs through the loader in common/ (not released; the anonymised trial table is ../data/userstudy/glmm_trials.csv). Results: ../expected/userstudy_rank_stability/. |
diffusart_retrain/ |
Diffusart baseline rows (Tables II/III) and the supplementary "additional training comparisons" (Diffusart-retrain) | code/: our complete Diffusart re-implementation (training loop, deterministic-hint loader data/data_load_det.py, train_det_{scribble,dot}.py, infer_det_proposed.py, evaluation; README inside); R3-2_README_original.md. Checkpoints: release v1.2. Results: ../expected/additional_training/. |
coldiff_finetune/ |
Table II/III "ColDiff (200ep)", supplementary "(7ep)" | run_coldiff_finetune_eval.sh, coldiff_v{1,2}_inference_official_fixed*.py (deterministic-hint inference wrappers around the official ColorizeDiffusion code; CC BY-NC-SA 4.0, see LICENSE in that directory), EXPERIMENT_coldiff_finetune.md (protocol notes). |
Conventions shared by all launchers: hint ratios 0.00 0.01 0.03 0.05 0.10 0.25 0.50 1.00; sketch_type 0 / 1 / 2 =
sketch simplification / XDoG / SketchKeras in the dataset loader; outputs land in
<result_path>/dataset_name/BrownianBridge_<hint>_illust/sample_to_eval/illust/<hint>/<sketch_type>/<ratio>/200/<id>.image.png;
inference seed 1234; the evaluation resizes to 256 × 256 and the Hint-AUC is the trapezoidal integral over α ∈ [0, 1].
The split lists referenced by the launchers (configs/illust/{train,valid,test}_paper.txt) are identical to
../data/splits/{train,valid,test}.txt (see SPLIT_LISTS.md).