visionary

Visionary is a desktop application for building image training datasets entirely on your own machine. It guides you through a workflow: ingest images from a folder, cull the ones you don't want, group similar shots together, tag them, write captions, and export a clean dataset ready for AI model training. The developer describes it as sitting between a file browser like Adobe Bridge and a training tool, focused on the curation step rather than editing or training. It is currently in alpha, runs from source, and is built primarily for macOS on Apple Silicon, with experimental Linux support and untested Windows compatibility.

To install it, you need Python 3.11 or newer. The recommended approach uses a tool called uv or pipx to create an isolated install that puts a visionary command on your system. You then run that command and open a folder of images. Visionary scans your files into a local SQLite index stored in a hidden .visionary folder alongside your images. It computes thumbnails and visual features in the background while you browse. Your original image files are never modified.

The grid view offers six arrangement modes: by visual structure, color, person, resolution, filename, or file size. You select images with clicks and keyboard shortcuts familiar from file browsers, then delete unwanted shots to the system Trash. A deduplication feature finds near-duplicate groups and marks all but the highest-resolution copy for review, while exact duplicates can be removed automatically. On macOS, a face detection tool groups images by person using on-device recognition, and nothing leaves your machine.

Captioning is a central feature. Taggers append keyword-style tags to each image's caption text without overwriting what exists. Vision language models write descriptive prose, with a built-in offline describer that needs no download and optional larger models like JoyCaption Alpha Two, Moondream 2, and Qwen2.5-VL. You can combine taggers and captioners for hybrid captions. The training section inserts a trigger word into captions, and export packages everything as a flat folder with .txt sidecars or a metadata.jsonl file, converting formats like HEIC and RAW to JPEG so training tools can read them. You can also push the packaged dataset to a remote machine via SSH.

Captions and tags are stored in .txt files next to each image, keeping datasets portable. Visionary does not watch the filesystem for changes; you press a refresh button to reconcile any external modifications. The full README is longer than what was shown.