LAXMAYDAY/anime_background_edit_data

Anime Background Edit Pairs

This dataset contains 20,221 paired anime background edits. Each row uses one image as ref1, an English editing instruction as edit_instruction, and another image from the same background group as target.

Fields

• ref1: source image to edit.
• target: desired edited image.
• edit_instruction: executable English instruction transforming ref1 into target.
• prompt_path: matching standalone prompt text file.
• pair_id and group_id: source grouping identifiers.
• annotation_model and annotation_source: annotation provenance.
• difference_summary, changes, and preserve: structured comparison data.
• confidence: model confidence from 0 to 1.
• invalid_pair: annotation-time flag for unrelated, unusable, or effectively unchanged pairs.

The dataset retains 183 completed rows marked invalid_pair=true so downstream users can choose whether to keep identity/no-op examples or filter them. The final 9 pairs that did not pass annotation validation are listed in abandoned_pairs.jsonl and are not part of the train split.

Annotation models

• doubao-seed-2-0-mini-260428: 18,131
• gpt-5.5: 2,090

Hugging Face layout

train/
  ref1/
  prompt/
  target/
  metadata.jsonl
abandoned_pairs.jsonl
PREPARATION_SUMMARY.json
README.md

metadata.jsonl uses ref1_file_name and target_file_name, following the multi-image ImageFolder convention used by LAXMAYDAY/anime_resolution_upscale_data, so the Hugging Face Dataset Viewer can expose separate ref1 and target image columns.

Images are WebP files converted at quality 95. Source and target retain their original dimensions and composition. Multiple edit pairs may reuse the same underlying image when a background group contains more than two variants.