DATASET WORKBENCH
Illustrious style dataset builder
Local workspace
FROM RAW IMAGES TO TRAINING-READY FILES
Build a clean, consistent style set.
Cull hard, tag what is visible, and export matching captions without leaving your browser.
Prepended to every exported caption.
STEP 01
Review the batch
Keep clear, varied images. Discard duplicates, low-resolution work, and pieces that do not represent the style.
Drop a batch here
PNG, JPG, WEBP, or GIF · keep originals at their working resolution
STEP 02
Tag what is actually there
Booru tags describe visible content. The natural caption adds useful context; it does not replace the tags.
Keep tag vocabulary consistent across the set.
Tags ≠ styleUse tags for subject, pose, clothing, setting, lighting, and composition—not your style name.
Repeat sparinglyFor a small set, repeat a useful tag in roughly 1–3% of captions, not everywhere.
Consistency winsUse the same spelling and granularity. Remove near-duplicates before export.
STEP 03
Preview the training folder
The download contains one image and one matching .txt file per kept image.
Images0
Caption files0
Triggermy_style
Caption format
Trigger first, then comma-separated booru tags, then a natural-language caption.
my_style, 1girl, portrait, ...
Keep at least one image to preview an export.