Home

Bitcrush Studio

Bitcrush Studio is a desktop application for building, captioning, cleaning, reviewing, and preparing large image datasets for training AI image models.

Studio organises your work around datasets — folders of images that Studio indexes so it can track their captions, tags, quality scores, and other information as you work. Most tools operate on your active dataset and the folder currently open in Gallery View, so you rarely need to type a path by hand.

This wiki documents each of Studio's tools. The pages below are grouped by the stage of dataset preparation they belong to.

Building a dataset

• Builder Tool — create and expand datasets by importing from Hugging Face, downloading from supported websites, or carving a focused dataset out of an existing one.
• Sort Tool — file loose or unsorted images into concept folders, with suggested matches to speed the process up.

Captioning and tagging

• Captioning — write, generate, and maintain captions across your dataset, including natural-language captions from a vision-language model and automatic tag-style captions.
• Tagger Tool — apply structured, Danbooru-style tags automatically using multi-label tagger models.

Cleaning and curating

• Gallery View — the central hub for browsing, filtering, sorting, and managing every image in a dataset.
• Quality Tool — score images automatically or by manual review, then cull the weakest from your dataset.
• Duplicate Tool — find and remove duplicate and near-duplicate images.

Preparing for training

• Masking Tool — draw or generate masks for masked diffusion training.

Getting help

If you run into a problem or would like to suggest a feature, please raise it in the Issues tab.