PixlStash 1.8 adds a justified grid. Thumbnails keep their own aspect ratio and each row is filled to the full width of the grid, instead of every picture being cropped to a square. A single thumbnail-size control, from Tiny to Huge, replaces the columns slider and sets the size in both layouts. The square grid is still available.
Rows are filled to the grid width at every scroll position, and the date pill shows the date range of the pictures currently on screen. Selection, rating, and keyboard navigation work the same as in the square grid.
The layout is set in Settings under Appearance. The same pane sets the sidebar thumbnail size and switches the sidebar between full width and a narrow icon dock. The grid thumbnail size is in the toolbar, next to the sort and filter controls.
Delete forever in the Scrapheap now removes the files from disk. It lists what will be deleted and asks you to type to confirm. Auto-empty is off by default: nothing is removed on a timer until you set a retention window of 30, 60, 90, or 120 days in Settings. Installs upgrading from 1.7 stay off until you set one. Originals in reference folders are never auto-purged.
Imports stream to the server and complete there. Closing the tab or navigating away does not cancel them. A running import appears in the task manager with progress, and can be aborted from there.
The stats panel has a new chart that plots your star ratings against the Smart Score. Cells off the diagonal are pictures where the two differ: rated high and scored low, or the reverse. Clicking a cell opens those pictures in the grid.
The whole library is re-scored in the background the first time you open 1.8, so the grid may re-order compared to 1.7. Two things changed. The built-in reference points used to score a library with few ratings of its own had stopped loading and now load again, so if you have rated only a handful of pictures your scores will change more than most. The scoring weights are also rebalanced so the top of the 1 to 5 range is reachable. The recompute runs in small batches and does not re-run any AI models. Originals and snapshots are not modified.
Every API route now goes through a single deny-by-default authorization gate that enforces object access from a declared policy. A route with no declared policy is refused rather than served, so a share link or API token can only reach what it was granted. This release also fixes snapshot restore hard-deleting files that were added after the snapshot was taken, and stops deletions removing pictures in locked sets.
Download the desktop app for Windows, macOS, or Linux. Or keep using pip, Docker, or the source build.
PixlStash 1.7 adds a tag review queue. It finds the tags most likely to be wrong or missing in your library and walks you through them picture by picture. Each card shows why the picture was flagged and asks one question: should this picture have the tag or not. Answer with a click or a single key, and the tagger learns from your decisions.
The tag health view ranks every tag in your library by how much a review is likely to fix. It estimates how many pictures have the tag wrongly, how many are missing it, and where the tagger disagrees with your past decisions. Start a review straight from the row, and filter by project, set, or character to keep the session focused.
Lock a picture set to freeze its pictures' tags, descriptions, and scores. Locked sets are read-only until you unlock them, and a lock badge shows wherever the pictures appear. Useful for eval sets and training data that must not drift after you have curated them.
A new Segment action detects objects and draws labelled boxes in the grid and the picture view. Export the boxes as COCO or Ideogram JSON to use them in your own training and annotation pipelines.
The whole app has been restyled on a design token system, with consistent type, colour, spacing, and icons. Dialogs, menus, and settings follow one visual language, in both light and dark themes.
Smart Score is now calibrated and precision aware. It weighs each defect tag by how confident the tagger is, groups related defects so correlated flaws do not count twice, and adds three new full-image checks for compression blockiness, noise, and watermarks. Your whole library is re-scored in the background the first time you open 1.7, so the grid may re-order compared to 1.6. That is expected. Your originals and snapshots are untouched.
PixlStash 1.6 introduces the PixlStash Desktop app for Windows, macOS, and Linux: one download that wraps the local server in a native window. It runs offline out of the box and opens straight into your library. No terminal, no Docker, no login screen.
Already running PixlStash from pip, Docker, or the Windows installer? The desktop app picks up your existing library, so nothing moves and nothing is lost.
The app ships with a built-in runtime that works on any machine: CPU on Windows and Linux, and Apple Silicon GPU acceleration (Metal) on macOS out of the box. On Windows and Linux, if you have a discrete GPU, add acceleration on demand from the Backend settings and choose from NVIDIA CUDA or experimental AMD ROCm libraries download straight from PyPI into a private overlay, and the bundled runtime is never touched. The same panel turns on optional remote access so other devices on your network can open the library in a browser in the same way as PixlStash server.
Everything is local. The bundled runtime means the app works with no internet connection, and it opens directly into your library with no login prompt because it talks to its own private local server. Close the window and PixlStash keeps running in the system tray, so your library, and any remote access you enabled, stays a click away. The pip, Docker, and source installs are all still fully supported; the desktop app is simply the easiest way to run PixlStash on your own machine.
PixlStash 1.5 puts a safety net under all your organising work. Automatic, compressed snapshots capture your library's metadata: tags, scores, faces, descriptions, and how everything is organised, so any mistake is a few clicks from undone. Roll all your metadata back, or restore just the pictures you select.
(Your image files are never touched.)
A full restore rolls all your metadata back to a point in time, but you don't always want that. Select any pictures and restore just their metadata from a snapshot, leaving everything else exactly as it is.
A breadcrumb trail above the grid always shows where you are: project, character, or set. This allows you to jump back up the hierarchy with a single click. No more guessing how you reached a view or navigating back from scratch.
Reorganise on the fly: grab a character or picture set in the sidebar and drop it onto another project. Their pictures are re-assigned for you. The move is reference-aware, so a picture that still belongs elsewhere stays put. Reshaping large libraries is now a drag away.
A snapshot copies only your metadata database, never the image files, which would be wildly impractical to duplicate over and over. That database is compressed with zstd for roughly a 3× saving, yet still keeps the expensive parts, CLIP embeddings and face data, so a restore comes back fully populated with no re-tagging or re-embedding pass afterwards. Retention follows a true grandfather-father-son policy: the most recent 7 daily, 4 weekly, and 12 monthly snapshots are kept and the oldest are pruned for you.
Everything the snapshots panel does is a plain REST call, so you
can script metadata snapshots and restores straight into your own
pipelines. Authenticate with an owner token and talk to the
endpoints under /api/v1.
Authorization: Bearer <token>
List every restore point, newest first.
curl -H "Authorization: Bearer $TOKEN" \
https://your-server/api/v1/snapshots
[
{
"id": 42,
"kind": "DAILY",
"label": null,
"created_at": "2026-05-31T06:42:37Z",
"picture_count": 1030,
"byte_size": 4523754,
"is_compatible": true
}
]
Take a manual snapshot right now, optionally labelled.
curl -X POST https://your-server/api/v1/snapshots \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"label": "before big cleanup"}'
{ "id": 51, "kind": "MANUAL", "label": "before big cleanup", "picture_count": 1030 }
Dry-run a full restore. It counts only what actually changes.
{
"summary": {
"pictures_to_revert": 12,
"pictures_to_recreate": 3,
"pictures_to_delete": 1,
"pictures_unchanged": 1014,
"missing_files": 0
},
"warnings": []
}
Restore all metadata to this snapshot. A safety snapshot is taken first.
curl -X POST https://your-server/api/v1/snapshots/42/restore \
-H "Authorization: Bearer $TOKEN"
{ "snapshot_id": 42, "resource_type": "full", "missing_files_count": 0, "errors": [] }
Selective restore. Bring back only the resources you list.
curl -X POST https://your-server/api/v1/snapshots/42/restore/batch \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"resources": [{"type": "picture", "id": 8123},
{"type": "picture", "id": 8124}]}'
# pip install requests import requests r = requests.post( "https://your-server/api/v1/snapshots/42/restore/batch", headers={"Authorization": f"Bearer {token}"}, json={"resources": [{"type": "picture", "id": 8123}]}, ) print(r.json()) # {'snapshot_id': 42, 'upserted_count': 1, 'errors': []}
{ "snapshot_id": 42, "upserted_count": 2, "errors": [] }
PixlStash 1.4 adds powerful new ways to find pictures. Reverse image search is now a right-click away in the grid and with a full API, wired straight into your generation workflows with the new ComfyUI-PixlStash nodes.
The
ComfyUI-PixlStash
nodes expose PixlStash search inside ComfyUI. Drop in a
reference image and the
PixlStash Likeness Search node returns the
closest matches from your library. Switch between
face_search and picture_likeness modes
to match on a person or on an overall picture. Store the
pictures straight back into your library with the
Picture Saver node.
Describe what you're looking for in plain language: "man searching for photos" and the PixlStash Semantic Search node returns matching pictures ranked by relevance, with a similarity threshold and result limit you control.
Batch multiple characters, sets, and individual pictures into a
single likeness search and combine them with
mean or min. Surface the pictures
closest to all of your inputs at once, not just one. Old man
drinking beer by a river? No problem!
The grid keeps getting quicker. 1.4 builds on the 1.3 loading overhaul with further optimisations, so even very large libraries feel more responsive as you scroll, filter, and search.
The image grid now loads visually very quickly and fills in details asynchronously. Even at 30 000+ images it feels snappy. Plus full JoyCaption support, selectable tag and description engines, and a URL that remembers where you were.
The settings page now lets you choose which plugin handles automatic tagging and which generates image descriptions. Each engine has its own parameter controls, so you can tweak precision, prompt style, and thresholds without leaving the settings panel.
Every view, whether grid, character, picture set, or project, has its own URL. Bookmark it, share it with a teammate, or simply refresh and land exactly where you left off. No more navigating back from scratch after an accidental reload.
Right-click any image (or use the selection toolbar) to re-tag or regenerate a description using the engine of your choice. A hover submenu lists every available model so you can pick the right one for the job without touching the settings page.
The image grid renders the visual content immediately and fills in metadata such as scores, tags, and faces asynchronously in the background. The result is a dramatically snappier feel: thumbnails appear before the first server response has even finished, and the grid stays responsive while data loads in. This is the single biggest day-to-day improvement in 1.3.
The biggest visual overhaul since launch: a redesigned sidebar, new toolbar, and context menus throughout. Plus, share pictures, sets, characters and projects with a single link.
Share individual pictures, sets, characters, and projects with anyone using a read-only link. No account is required for viewers. Tokens are scoped, revocable, and never expose your full library. Ideal for sharing work-in-progress with clients or collaborators. Optionally add your own watermarks to shared pictures for branding.
Recipients see a clean, read-only view of exactly what you shared. No login, no noise, just the content you chose to share.
The toolbar has been redesigned for clarity. Search, filters, and sort controls are better organised, and the new layout scales more naturally on smaller screens.
Right-click anywhere in the grid or sidebar to access actions for the item under the cursor. Rate, tag, assign, move, share, or delete without leaving your current view.
Give each picture set its own coloured icon instead of relying on the automatic thumbnail. Instantly recognise your sets at a glance, no matter how similar the contents look.
People, sets, projects, and folders are easier to navigate. The new sidebar is more compact, loads faster, and shows more at a glance.
Tagging and face extraction are significantly faster thanks to bulk database operations and smarter grid update scheduling. Watch real-time throughput in the statistics sidebar task manager.
PixlStash now checks for known security issues in the ChangeLog and alerts you directly in the UI when a fix is available along with the security level and affected versions. Stay informed and keep your installation secure with ease.
This release adds reference folder support, letting you keep your existing folder structure while still benefiting from PixlStash's powerful organisation and metadata features.
Get useful statistics about your picture sets. See tag distribution and tag prediction confidence. A powerful way to understand your training sets.
See statistics about your pictures including size distribution and score distribution. Filter image views by clicking on elements in the sidebar. Quickly find outliers to review or delete.
Hold Ctrl and click names in the People or Sets sidebar to build a multi-selection, then choose how to combine them using the toolbar at the bottom of the grid. Union shows all pictures from any selected person or set. Overlap finds pictures where every selected person appears together. Difference subtracts one set from another, useful for finding pictures unique to a base set. Unique (XOR) returns pictures that belong to exactly one of the selections, letting you quickly spot what each person or set contributes exclusively.
