Sampling & Repeats

Repeats

Set to 0 for a single pass. Each repeat adds another sweep through the dataset.

Probability

Relative sampling weight when mixing datasets.

Enable "I Know What I'm Doing" in Advanced settings to adjust probability.

Embed Options

Write Batch Size

Controls embed cache flush size.

Aspect Ratio Limits

Min Aspect Ratio

Skip images with aspect ratio below this value.

Max Aspect Ratio

Skip images with aspect ratio above this value.

Discovery Overrides

Disable discovery only after embeds and caches are fully generated.

Caption Filtering

Caption Filter List

Provide a path or inline JSON/line list to apply custom filtering.

Embed Dataset Links

Text Embeds Dataset

Leaving this empty keeps caption embeds with this dataset using its cache directory.

Image Embeds Dataset

None keeps image embeds in this dataset's storage backend; selecting another dataset reuses that dataset's embed cache location.

Cache & Processing Options

Hash cached filenames

Creates shorter, hashed names for VAE cache files to avoid filesystem path length limits.

Preserve backend cache

Keeps the file list cache between epochs. Essential for large datasets on AWS S3 or slow storage.

Regularisation data

Uses the base model's predictions on this dataset to prevent the model from drifting outside the intended class token. Requires another non-regularisation image/video dataset.