Just some of the many things PixlStash can do
Semantic search, face recognition, tagging, captioning, scoring and a review queue that checks the models' work. All of it runs on your machine.
Describe what you are looking for and get pictures back, tagged or not. CLIP embeddings, computed locally.
Faces are detected, embedded and grouped into characters you name. Search the library by uploading one face.
PixlStash Tagger, WD14, Florence-2 or JoyCaption, chosen per request rather than per install. Re-tag any selection from the context menu.
A ranked queue of the tags the models most likely got wrong, plus a health board per tag. Auto-tags become tags you can trust.
Smart score sorts a large set by what is worth looking at, with a calibrated penalty for malformed anatomy and render defects.
Generate from a selection and the results come back tagged, with their people, sets and projects already attached.
Point PixlStash at a folder and it reads it where it sits, turning the folder names you already chose into projects, people and sets — without moving, renaming or copying a file.
One folder picker, and the folder answers. A library you already made is added as it is, with its tags, scores and people. A folder of pictures is counted and offered. An empty folder starts a fresh library. A folder another library already covers is refused, and the refusal names that library.
PixlStash proposes what each folder level is — project, set, person, tag or just a folder — from where the pictures sit, how the names are shaped, when they were taken and whose faces are in them. Take a level from its band or override a single row.
The review screen names every project, set and person your folders are about to become, and says plainly that no file is copied, moved or renamed. Organise later brings the pictures in and leaves the naming for another day; an interrupted import resumes.
Give a library a layout — Project then Person or Set — and PixlStash writes new pictures where it says. Choosing one moves nothing that is already there, and renaming a project, person or set renames their folder without moving a file.
A picture moves only when its folder stops being true, so adding a second project or person moves nothing. Where a picture is filed truthfully but not where the layout would put it today, the move is offered, counted first, and undone as one batch.
Export the pictures you selected into a folder on the machine running PixlStash, which then opens in your file manager. Captions, resolution, sidecars and face crops come out as the zip export produces them, without the round trip through downloads.
Keep separate picture libraries in one installation, catalogue the models already on your disks, and run the housekeeping from a terminal.
Register more than one library and keep one open at a time. Pictures, tags, scores and snapshots stay with their library, while your account, your preferences and the model shelf stay with the installation. Switch from Settings → Libraries.
API tokens and share links belong to the library that created them. They go quiet while another library is open and work again when you switch back, and Settings warns you before you switch away from a library with live share links.
Register the folders holding your checkpoints and LoRAs and see what is in them: kind, base model, size, and every folder a copy sits in. Group and sort the shelf, move a model between folders, or delete one to your Trash.
Read an ai-toolkit run where it was written, then import it with its sample images. Versions of one subject group behind a single row, so a character you have trained four times is one entry with its history under it.
pixlstash-cli manages libraries, writes and restores backups, and installs plugins. Every command documents itself, including the exit codes a script needs, so a nightly backup is one cron line. The desktop app can run it without being open.
The desktop app checks at start-up whether another account on the computer can write to your library folder, and offers to tighten the permissions. It changes permissions only, and leaves your own access alone.
Reverse everyday edits across your library, review duplicate candidates, and keep people and sets organised across projects.
Reverse soft-deletion, tagging, description edits, image assignments, moves, and duplicate decisions with standard keyboard shortcuts or toolbar controls. Open the history to inspect recent changes or step back through several at once. It is a dependable safety net for professional library work, not just the last button you clicked.
Work through likely duplicate groups in a dedicated queue. Compare candidates, choose the cover, exclude a mismatch, stack the group, or keep it separate—and undo the decision if it was not right. Optional near-duplicate tiers and a similarity threshold put you in control of how broadly PixlStash searches.
Filter a view to stacked or unstacked pictures. Keep cover only previews a stack cleanup, keeps its cover, and sends the other copies to the recoverable Scrapheap without deleting originals from disk.
Add one character or picture set to several projects at the same time. Its pictures appear in each project, while removing one membership leaves every other project alone.
Start from a person and rank the library by face likeness. Adjust the match, review the result, and assign all found pictures or only your current selection in one action.
Update checks and anonymous install telemetry have their own independent controls. Both are off until you enable them, and the Privacy pane states what is and is not sent.
See the checkpoints and LoRAs in the folders you register, grouped by kind, base model, folder or feature.
Switch between registered libraries from Settings. Each one keeps its own pictures, tags, scores and snapshots.
Inspect named changes, undo one or several steps, and redo them across metadata edits, assignments, moves, and stack decisions.
Compare likely copies, choose a cover, stack the group, keep it separate, and undo the decision when it was not right.
Keep one person or picture set in several projects without duplicating it or moving it back and forth.
Pretty image view with scoring, stacking and sorting.
Thumbnails can keep their own aspect ratio and fill each row to the full grid width, or stay as squares. One size control in the toolbar, from Tiny to Huge, sets the thumbnail size in both layouts.
Organise your projects efficiently with intuitive controls and additional metadata (here in dark theme)
Easily find pictures containing particular characters.
Right-click any picture to run a reverse image search for visually similar pictures, or find other pictures containing the same face.
A review queue finds the tags most likely to be wrong or missing and asks one question per picture. Answer with a click or a single key to clean up your tags.
Detect the objects in a picture and draw labelled boxes over them. Export the boxes as COCO or Ideogram JSON for training and annotation pipelines.
Automatic, compressed snapshots of your metadata, on a daily, weekly, and monthly schedule. Roll back to any restore point in a few clicks; a safety snapshot is taken first.
Don't want to roll back everything? Restore just the pictures you select. Anything not selected is left untouched.
Select images, tag or assign to sets or characters and perform batch actions.
Run your own ComfyUI workflows directly from PixlStash with progress tracking.
See the ComfyUI nodes & workflows →Review and edit automatic descriptions and tags.
Detect and assign faces to people in your photos.
Tag with a keyboard shortcut and let PixlStash suggest tags as you type.
Manage and organize people in your dataset efficiently.
Lots of convenient keyboard shortcuts for navigation, selection, scoring, tagging, and more.
Keep your existing folder structure while still benefiting from PixlStash's organisation and metadata features. Manage reference and import folders directly from the UI.
Get a rapid overview of your picture sets, tag distribution, tag prediction confidence, and picture statistics at a glance.
Ctrl-click multiple characters or picture sets, then combine them with Union, Overlap, Difference, or Unique (XOR) operations.
Extend PixlStash's functionality with custom Python plugins for image processing.
Share pictures, sets, characters, and projects with a read-only link with no account required for viewers. Tokens are scoped and revocable. Optionally add your own watermarks for branding.
Right-click anywhere in the grid or sidebar to rate, tag, assign, move, share, or delete without leaving your current view.
Give each picture set its own coloured icon for instant recognition, no matter how similar the contents look.
The rest of the toolkit. The AI features have their own page.
Extend PixlStash's functionality with custom Python plugins for image processing.
Navigate and control PixlStash efficiently using standard keyboard shortcuts for selection, scoring, deletion, etc.
Deleted pictures go to the scrapheap first and can be put back. PixlStash can empty it after 30, 60, 90, or 120 days, or leave it alone if the retention is set to Never. Auto-empty is off by default, and deleting for good asks you to type to confirm.
Imports stream to the server and complete there, so closing the tab or navigating away does not cancel them. Progress shows in the task manager, and a running import can be aborted from there.
PixlStash checks for known security issues and alerts you in the UI when a fix is available, along with the severity and affected versions.
Each picture moves to the next model as soon as its own rows land rather than waiting for the whole batch, models stay resident for a pass, decoding and resizing happen off the GPU worker, and the VRAM budget follows your card. Watch real-time progress in the statistics sidebar.