CoolFace
Datasetpublic

jjobear/collage-layout-dataset

Collage Layout Synthetic Dataset Synthetic photo-collage layouts for layout-quality analysis & correction, built on a six-ingredient framework (Format, Photos, Visual Weight, Hierarchy, Readability, Harmony). Corrector-not-generator: every collage carries a naive (v1_center) and a corrected (fit) placement, so a model can learn the correction. Faces are synthetically replaced (privacy-safe). How to load from datasets import load_dataset ds =… See the full description on the dataset page: https://huggingface.co/datasets/jjobear/collage-layout-dataset.

sourceHugging Facecc-by-nc-4.0updated 2mo agoView on Hugging Face
0likes53downloads
Dataset Card

Collage Layout Synthetic Dataset

Synthetic photo-collage layouts for layout-quality analysis & correction, built on a six-ingredient framework (Format, Photos, Visual Weight, Hierarchy, Readability, Harmony). Corrector-not-generator: every collage carries a naive (v1_center) and a corrected (fit) placement, so a model can learn the correction. Faces are synthetically replaced (privacy-safe).

How to load

python
from datasets import load_dataset
ds = load_dataset("USERNAME/collage-layout-dataset")   # dict with train/val/test/gold
ds["gold"][0]        # -> {"image": <PIL>, "collage_id": ..., "gold_overall": ..., ...}

Schema (long format)

Each row is one rendered collage image plus its annotation. Because every collage is rendered under two placement strategies, a collage appears as two rows distinguished by the strategy column (v1_center = naive, fit = corrected). To recover the corrector-not-generator pair, group rows by collage_id and read the two strategy values. (This is the "long" equivalent of a wide poster-layout schema that would put both renders as side-by-side columns.)

Columns

columnmeaning
imagethe rendered collage (PIL image)
collage_idunique id (run/cNNNN); the two strategies share it
strategyv1_center (naive) or fit (saliency-aware corrected)
splittrain / val / test / gold (album-level, leakage-clean)
albumsource album the photos came from
cross_albumtrue = intentional incoherent "don't-belong" negative
scene_typesscene tags for the photos
canvas_name, n_slots, margin, gutter, split_treelayout format
slots_jsonslot rectangles (x,y,w,h)
assignment_jsonordered photo_ids placed into the slots
placements_jsonper-photo crop/scale/strategy + reject reasons
all_feasible, n_infeasible, placement_flagsplacement feasibility
balance_offset, balance_flagsclassical centroid-balance metric (NOT a quality predictor — see below)
gold_*5-designer consensus, gold split only: gold_overall, gold_balance, gold_hierarchy, gold_readability, gold_harmony, gold_ship_rate, gold_overall_std, gold_n_raters. In train/val/test these are not rated: sentinel -1.0 (and gold_n_raters = 0). Filter with gold_n_raters > 0.
run_mode, face_backend, saliency_backend, weight_roi, weight_saliencyprovenance

Splits (album-level, leakage-clean)

Whole albums are held out, so train/val/test/gold never share photos. Rendered-image rows in this release: train 120, val 80, test 80, gold 400. The full recipe-level table (every collage, with or without a rendered image) is in all_collages.jsonl in the pipeline repo.

Validated quality signal

assignment_confidence (photo→slot fit) predicts the 5-designer consensus on every axis (~+0.35 original study; +0.57–0.58 on a held-out gold set with 5 new raters), at or above the ceiling human inter-rater agreement allows. balance_offset does not predict quality (~0 to −0.2). See the poster / repo.

Known limitations

  • —Quality is moderately subjective (designers agree ~+0.33–0.45 overall; hierarchy/readability weakest).
  • —assignment_confidence is blind to semantic/color coherence — it does not flag cross-album "don't-belong" negatives. A coherence metric is future work.
  • —Single-photographer album pool; synthetic faces. cross_album=true rows are intentional incoherent negatives.

Ethics

Faces are AI-generated replacements; original identities never enter the release. License CC-BY-NC-4.0 (recheck any downstream model licenses for commercial use).