XIAOHAOYANG/figure4-visualization-dataset

Ordinal-Reference Visualization Dataset

Created: 2026-07-24 UTC
Status: final frozen 50-case package

This package is a hand-curated qualitative dataset for horizontal, same-case image-editing comparisons. Each image contains repeated instances of one subject class arranged in one or two visually legible horizontal rows. The cases deliberately span food, plants, household and laboratory objects, toys, vehicles, wildlife, and people so that ordinal grounding is tested across more than a single visual domain.

The final freeze contains 50 source images, 268 ordered subject instances, and 100 paired A/B model jobs. Forty images have one row and ten have two rows. The preferred count is 4--8 instances; the sole count exception is the strict 2x5 PCB case with 10 instances. The displayed long edge ranges from 1,920 to 8,256 pixels. This is a qualitative visualization set, not a statistically representative benchmark.

Selection and freeze policy

A case is eligible only when all of the following hold:

• exactly one repeated subject class is being indexed;
• all indexed subjects are complete enough to count and edit;
• the subjects form one or two visually legible horizontal rows;
• ordinals are unambiguous when read from the viewer's left to right within the named row;
• the preferred count is 4--8, with a hard maximum of 10 and an explicit note for any case above 8;
• source resolution and visible detail are suitable for a paper-scale comparison;
• the source page, creator, rights statement, processing history, and local hash are retained; and
• two nonidentical A/B targets and one shared edit operation are frozen before any editing-model output is inspected.

Every final source image must be reviewed at full composition, with zoom as needed to resolve count, boundaries, and small details. Contact sheets and subagent proposals are discovery aids, not substitutes for that review. The main reviewer also checks every numbered instance-order overlay and every A/B SAM2 overlay before a case is marked final. Images with clipped outside subjects, hidden extra instances, ambiguous depth clusters, or row geometry that cannot support a stable ordinal are kept outside the release.

The paired-prompt invariant

Each case stores exactly one prompt_template with exactly one {target_expression} placeholder. A and B are generated by substituting different frozen ordinal expressions into that same template. Nothing else may change between the pair:

• the edit action and component;
• the requested color, material, or state;
• the scope and preservation clause;
• the source pixels and input resolution; and
• the seed policy and output postprocessing used for a given model.

Only target_expression and its corresponding target_instance_id differ. For example:

Template: Repaint only {target_expression} vivid cobalt blue, preserving ... unchanged.
A:       Repaint only the second motorcycle from the left vivid cobalt blue, preserving ... unchanged.
B:       Repaint only the third motorcycle from the left vivid cobalt blue, preserving ... unchanged.

Freeze both targets before viewing any model output. Do not select whichever target makes a preferred method look better. This counterfactual pairing is what makes a wrong-instance edit interpretable rather than confounded with edit difficulty.

Public model input versus evaluator-private data

Ordinary editing baselines receive only the source image and complete natural-language prompt. All parsed grounding and spatial supervision stay evaluator-private.

Surface	Permitted fields or artifacts	Rule
Public baseline jobs	case_id, pair, source_image, prompt_full	Give the model only source_image and prompt_full; the IDs are orchestration labels.
Evaluator-private metadata	target_expression, target_instance_id, instance_order, row labels, boxes, point prompts, edit_spec	Never expose these to an ordinary baseline.
Evaluator-private visual assets	target boxes, isolated masks, mask overlays, target/distractor crops	Use only for checking, scoring, and the explicitly labeled reference/target column of a figure.

metadata/baseline_jobs.jsonl is the safe job list. It must contain exactly four fields per row: case_id, pair, source_image, and prompt_full. Do not derive or attach a target crop, parsed locator, bounding box, coordinate, or mask for a baseline unless that model's published task definition explicitly requires spatial input and the comparison is labeled accordingly.

Directory layout

visualization_dataset/
  images/                         final model inputs
  metadata/
    case_specs.json               human-reviewed source of truth
    manifest.json                 expanded evaluator-private manifest
    baseline_jobs.jsonl           public-safe A/B model jobs
    prompt_pairs_private.jsonl    evaluator-private expanded targets
    source_inventory.jsonl        source, rights, processing, and hashes
    CATEGORY_INDEX.md             human-readable category-to-case index
    ORDER_AND_PROMPTS.md          evaluator-private order and A/B prompt index
    sam2_masks.jsonl              SAM2 audit rows
    sam2_run.json                 checkpoint/config/run record
    sam2_quality_audit.*          structural and heuristic mask audit
  masks/                          evaluator-private binary target masks
  mask_overlays/                  evaluator-private human-review overlays
  target_boxes/                   evaluator-private box/point previews
  previews/instance_order/        evaluator-private numbered-order overlays
  coordinate_grids/               annotation aids
  contact_sheets/                 visual audit pages
  scripts/                        deterministic build, mask, and validation tools
  review_rejects/                 rejected near-misses, excluded from the manifest

Do not copy evaluator-private directories into a baseline input bundle.

The final contact-sheet families are dataset_sources_01..05, instance_order_audit_01..09, sam2_overlays_01..10, sam2_binary_masks_01..10, and sam2_target_boxes_01..10 under contact_sheets/.

Mask generation and semantics

Evaluator masks are generated locally with SAM2.1 Hiera Large using the exact checkpoint:

/workspace/models_weight/sam/sam2.1_hiera_large.pt

and config:

configs/sam2.1/sam2.1_hiera_l.yaml

The standard mask denotes the visible target instance, not necessarily the smaller component named by the edit. A prompt may recolor a shirt, painted body panel, scarf, or liquid while the evaluator mask covers the visible person, vehicle, statue, or vessel. If an experiment requires an edit-region mask, create and label a separate artifact; never silently reinterpret the instance mask.

Three declared semantic sub-instances are intentionally narrower: carrots_5 masks the edible orange roots rather than foliage, tulip_blossoms_5 masks blossoms/petals rather than stems, and eggs_6 masks visible eggshell rather than the carton. Transparent glass, thin rotors/legs/handles, foliage, inter-person occlusion, and overlapping vehicles require especially careful overlay inspection. A SAM2 confidence score is not a replacement for visual boundary review.

Reproducible build and review

Run commands from visualization_dataset/:

python scripts/build_metadata.py --dataset-root .
python scripts/make_human_indexes.py --dataset-root .
python scripts/validate_dataset.py --dataset-root . --allow-missing-derived
python scripts/make_instance_order_overlays.py --dataset-root .
python scripts/generate_sam2_masks.py --dataset-root . --device cuda
python scripts/audit_sam2_quality.py --dataset-root .
python scripts/validate_dataset.py --dataset-root .

The first validation checks the source freeze before derived masks exist. After the numbered overlays are visually approved, generate SAM2 outputs and inspect every A/B overlay and isolated binary mask. The final validation must pass without --allow-missing-derived.

For fair model comparison, preserve each case's source resolution, use the same seed policy per model, record any model-required resize, and apply identical output resizing in the paper. A target/distractor zoom can help at five-column paper scale, but it must not replace the full output.

People, culture, marks, and synthetic edits

Recognizable-person cases are restricted to the frozen, benign garment-color operations. Do not change faces, identity, body shape, protected or sensitive traits, health status, rank, name, or behavior. Preserve uniforms, official identifiers, and scene context; label outputs as synthetic research edits and do not imply that any depicted person, employer, agency, or creator endorses a model or paper.

For people_environment_team_7, the official caption describes front 4 + back 3, while the benchmark freezes the visually readable pixel grouping as standing/back 5 + lower/front 2. The dataset row labels and ordinals are authoritative for model prompts and evaluation.

The Jizo-statue case receives only a neutral neck-scarf recolor. Preserve faces, bodies, inscriptions, offerings, and setting, and avoid disrespectful, defamatory, or endorsement-implying use. Toy figures, toy cars, motorcycles, vehicles, PCBs, sports equipment, clothing, signs, number plates, and other depicted objects may retain trademarks, product designs, or other non-copyright rights; source copyright permission does not clear those rights.

Rights and release boundary

See LICENSE_PROVENANCE.md and metadata/source_inventory.jsonl before publishing any pixels or generated outputs. Pexels and Unsplash use custom platform licenses rather than CC0. For a public benchmark, the conservative default is to distribute the annotations, source URLs, expected dimensions, and SHA-256 hashes, plus a fetch/verification procedure, rather than repackaging stock-platform originals. Recheck the current terms and the contemplated distribution form first.

The CC BY-SA case requires attribution, a license link, and a change notice, and may impose ShareAlike obligations on covered adaptations. U.S. government and DVIDS/USGS items retain privacy, publicity, trademark, official-mark, visual-identifier, and non-endorsement concerns. The Fort Riley Army image requires exact asset-status verification before raw redistribution.

No image, license label, public-domain statement, or automated check makes legal risk literally zero. This package records a good-faith provenance audit for research use; it is not legal advice, and the source page and license text control.