saaduddinM/spatial_edit_judge

Spatial Edit Judge Dataset

This dataset tests whether a visual judge can decide if a requested spatial camera edit was actually satisfied.

Each row contains a before image, an after image, a text instruction, and the ground-truth judge label. The judge should answer whether the after image correctly follows the instruction.

Dataset Structure

The Hugging Face viewer uses:

• data/train-00000-of-00001.parquet

The repository also keeps the raw exported files:

• images/{id}_before.*
• images/{id}_after.*
• instructions/{id}.txt
• metadata.jsonl
• metadata.csv
• skipped_non_invertible.jsonl

Columns

• before_image: source image before the edit.
• after_image: candidate post-edit image.
• instruction: camera edit command shown to the judge.
• type: one of correct, error_1, or error_2.
• error_kind: none, no_edit, or wrong_edit.
• expected_judge_label: success for correct edits and failure for error cases.
• is_successful_edit: boolean version of the expected judge label.
• dataset_origin: bytemorph or se.
• original_edit_command: the true edit command from the curated source dataset.
• wrong_edit_command: the opposite-direction command used for wrong-edit cases.
• actual_after_edit_command: the edit that the after image actually shows, when applicable.

Case Types

Total rows: 378

• correct: 126 rows. The after image is the true edited image for the instruction.
• error_1: 126 rows. No edit was done; the after image is a copy of the before image.
• error_2: 126 rows. The image pair shows the original edit, but the instruction is swapped to the opposite direction within the same edit type.

How It Was Created

This dataset was created from two manually curated camera-edit gold sets:

• ByteMorph camera gold set: 30 examples.
• SpatialEdit-Bench camera gold set: 101 examples.

Only examples with safe opposite-direction camera commands were used for this judge dataset. Five ByteMorph examples were skipped because they involved forward or backward camera motion and did not have a safe opposite under the chosen rules.

For every usable base example, three rows were created:

1. correct: before image plus true after image, using the original instruction.
2. error_1: before image plus the same before image as the after image, using the original instruction.
3. error_2: before image plus true after image, but the instruction was changed to the opposite camera direction.

Wrong-edit instructions were created only within the same edit type:

• zoom in <-> zoom out
• left <-> right
• up <-> down

No cross-type swaps were used. For example, zoom commands were only swapped with zoom commands, and pan/tilt commands were only swapped with their opposite direction.

The row order was shuffled with seed 42.

Distribution

Case distribution:

• correct: 126
• error_1: 126
• error_2: 126

Origin distribution:

• bytemorph: 75
• se: 303

Instruction distribution:

• move camera down and zoom in: 1
• move camera left: 13
• move camera left and point camera down: 2
• move camera right: 14
• move camera right and point camera up: 1
• move camera up and zoom out: 2
• pan camera left: 27
• pan camera right: 30
• point camera down: 58
• point camera up: 56
• zoom in: 87
• zoom out: 87

Intended Use

Use this dataset to evaluate models or rule-based judges that score whether an instruction-guided spatial image edit was successful.

The primary label is expected_judge_label:

• success: the edit matches the instruction.
• failure: the edit is missing or the instruction describes the opposite edit.