Caption & Metadata Strategy

Caption Strategy

How captions are stored and loaded.

Please select a caption source

Metadata Backend

How file metadata is indexed and accessed.

Trigger Words

Instance Prompt

Prepend instance prompt to captions

Enable to inject the instance prompt ahead of each caption when captions are present.

Instance prompt mode always uses this text without additional captions.

Caption Shuffling

Generate shuffled variants of tag-based captions for data augmentation during text embed caching.

Enable shuffling

Split On

Delimiter for splitting caption into tags.

Shuffle Count

Number of shuffled variants per caption.

Position Start

Keep first N tags fixed.

Seed

Optional seed for determinism.

Include original

Include unshuffled caption in variants.

Parquet Details

Parquet Path

Filename Column

Caption Column

Fallback Caption Column

Width Column

Height Column

Identifier includes extension

Assumes filenames already include their extension (recommended).