ImgMgr

A browser-based image manager built for large collections of AI-generated images (Stable Diffusion and compatible tools). Runs as a local Node.js server with a React frontend.

Features

• Virtualized tile grid — only the visible rows are rendered, so it stays fast with tens of thousands of images; pages stream in as you scroll
• Full-size lightbox with a rich metadata sidebar — separate Prompt / Negative prompt / Template sections that are collapsible and resizable, LoRA and wildcard highlighting, a resizable panel width, and the generation params surfaced first
• Folder tree — collapsible, with drag-and-drop move support, newest folders first
• Star ratings (0–5) with per-level counts in the filter bar; rate from the grid, the lightbox, or the 0–5 keys
• Keyboard-driven lightbox — arrow keys walk the entire result set (pages load as you go), 0–5 rate, F toggles 5★, Space advances, Del deletes and advances, I hides the info panel
• Slideshow — play the current view, your selection, or the whole library full screen, 1–20 s per image, looping back to the start
• Picks up where you left off — the grid remembers your scroll position per view across reloads
• Collapsible panels — hide the folder tree (B) and the metadata sidebar (I) to view images on the full screen
• Copy prompt — one click copies an image's positive prompt, negative prompt, or template, from a tile's hover button or the lightbox
• Trash & undo — deletes are recoverable: images move to a trash you can restore from, and an Undo appears right after deleting
• Offline-media friendly — if a drive or network share goes away, those images stay browsable (cached thumbnails, ratings, tags and metadata intact) and are just badged as offline; nothing is ever deleted without you asking
• Delete protection — starred images can't be deleted from any view
• Multi-select — select images, then move to a folder, delete, tag, or bulk-rate them
• Metadata facet filters — filter by Model, Sampler, or Steps dropdowns (extracted from the SD generation parameters), combinable with search and folders
• Find similar — from the lightbox, discover visually similar images using a DCT-based perceptual hash, with an adjustable strictness slider
• AI captions (VQA) — generate a natural-language caption for a single image or a whole selection via an SDNext server; captions are stored in the database and separately searchable
• Duplicate finder — two modes:
  • Exact — identical file content (MD5 hash)
  • Seed — same generation seed in the filename (might match non-similar images!)
• Live updates — watches today's output folder via SSE; a banner appears when new images arrive
• Metadata search — filter by filename, prompt text, or arbitrary EXIF key/value, with multi-term AND, exclusions (-term), and quoted phrases; target a single field with caption: / prompt: / name: prefixes
• Prompt phrases — the concepts recurring most in your prompts, counted and clickable, so you can browse the library instead of guessing search terms
• Reads SD metadata from PNG tEXt chunks and WebP/JPEG EXIF (positive prompt, negative prompt, seed, steps, etc.)

Requirements

• Node.js 22+ (uses --watch for dev mode)
• npm 11+ (bundled with Node 22; npm 12 also supported)
• Linux / macOS / WSL2

WSL2 note: If your images are on a DrvFs mount (e.g. /mnt/d/), the file watcher uses polling (usePolling: true) because inotify does not work across the Windows filesystem boundary. Set watchInterval in config.json to balance responsiveness vs CPU.

Setup

git clone https://github.com/tillerz/imgmgr
cd imgmgr
npm install

npm 12 note: npm 12 blocks package install scripts by default. The native dependencies (better-sqlite3, sharp, esbuild) are pre-approved via the allowScripts field in package.json, so npm install builds them automatically. After upgrading any of those packages, re-approve with npm install-scripts approve <pkg>.

Edit config.json to point at your image root:

{
  "port": 3000,
  "imageRoot": "/path/to/your/images",
  "thumbnailSize": 220,
  "thumbnailQuality": 82,
  "cacheDir": ".cache",
  "supportedExtensions": ["png", "webp", "jpg", "jpeg"],
  "scanOnStart": true,
  "watchForChanges": false,
  "watchInterval": 5000,
  "sdnextUrl": "http://127.0.0.1:7860",
  "captionModel": "Google Gemma 3 4B"
}

Key	Default	Description
imageRoot	~/sd/imgmgr	Root directory scanned for images
thumbnailSize	220	Max thumbnail dimension in pixels
thumbnailQuality	82	WebP thumbnail quality (1–100)
cacheDir	.cache	Directory for the SQLite DB and thumbnail cache (relative to project root)
supportedExtensions	["png","webp","jpg","jpeg"]	File types to index (JPEG/.jpg read SD params from EXIF UserComment)
scanOnStart	true	Run a full scan when the server starts
watchForChanges	false	Enable live file watching
watchInterval	5000	Polling interval in ms (used on DrvFs/WSL2 mounts)
trashDirName	trash	Sub-directory of cacheDir where deleted images are moved (recoverable)
sdnextUrl	http://127.0.0.1:7860	SDNext server used for AI captioning (its /sdapi/v1/vqa endpoint)
captionModel	Google Gemma 3 4B	VQA model name requested for captions (must exist on the SDNext server)
captionQuestion	describe the image	Prompt/question sent to the caption model
captionSystem	(see config.js)	System prompt sent to the caption model
captionTimeoutMs	300000	Give up on a caption request after this long (5 min). A cold model load is slow, but without a limit a wedged SDNext would hang the request forever
allowedOrigins	localhost/127.0.0.1 on port	Browser origins allowed to make state-changing requests. See Access control

Access control

imgmgr has no login — it assumes a local, trusted machine. But a browser will happily send requests to localhost on behalf of any page you have open, so the API restricts who may change things:

• Reads (GET/HEAD) are open.
• Writes (POST/PATCH/DELETE) are refused with 403 unless the request's Origin is in allowedOrigins, or there is no Origin at all (curl, scripts, the server itself).
• Filesystem paths in a request body (targetFolder when moving, path when creating a folder) must resolve inside imageRoot, so the API can't be used to write elsewhere on the machine.

By default allowedOrigins covers http://localhost:<port>, http://127.0.0.1:<port> and the IPv6 equivalent. If you open the UI on another address (a LAN hostname, say), add it:

{ "allowedOrigins": ["http://localhost:3000", "http://my-nas.local:3000"] }

Set it to ["*"] to disable the check entirely. Not recommended — that restores the state where any website could drive your API.

Running

Development (hot reload)

npm run dev

Starts the Express server with node --watch. Vite runs in middleware mode on the same port — no separate frontend process needed.

Open http://localhost:3000

Production

npm run build   # compiles the React app into dist/
npm start       # serves dist/ via Express

Usage

Browsing

• The folder tree on the left lists all indexed folders. Click a folder to filter. Click the ▸ arrow to expand/collapse subfolders.
• Use the search bar to filter by filename or positive-prompt text (see Search syntax below).
• Use the ★ buttons to filter by rating. Counts are shown next to each button, and they respect the other active filters.
  • The ≥ ★ / = ★ button flips between this rating and up and exactly this rating — handy when triaging, e.g. "show me only the 3-star ones".
  • ☆ N shows only unrated images.
• The Caption dropdown narrows to images that do or don't have an AI caption yet — the quickest way to see what's left to caption.
• The Model / Sampler / Steps dropdowns filter by the SD generation parameters (see Facet filters). They combine with search, folder, and each other.
• The sort dropdown supports date, name, and star rating in both directions.

Facet filters

imgmgr parses the generation parameters (the Steps: …, Sampler: …, CFG scale: …, Model: … line) out of each image's embedded metadata and stores them as discrete, filterable fields. The toolbar exposes Model, Sampler, and Steps dropdowns — each option shows how many images match — so you can quickly narrow to, say, everything made with a particular checkpoint and sampler. Selecting values in more than one dropdown ANDs them together, and the star-rating counts update to reflect the active facets.

Existing libraries are backfilled automatically on first launch after upgrading (a one-time pass over the database); newly scanned images get the fields during indexing.

Search syntax

The search bar matches against both the filename and the positive prompt. It supports multiple terms, exclusions, and phrases:

Type this	Matches images that…
sunset beach	contain both sunset and beach (space = AND)
sunset -beach	contain sunset but not beach
"close up"	contain the exact phrase close up
"close up" -blurry	contain the phrase close up but not blurry
castle + dragon - modern	contain castle and dragon, but not modern
caption:castle	have castle in the caption only
caption:"golden hour"	have the phrase golden hour in the caption
dog -caption:cartoon	match dog (filename/prompt) but exclude cartoon captions
phrase:"arabic eyeliner"	use that prompt concept, however it is spelled

• Space means AND — every included term must be present.
• - excludes a term. It works attached (-blurry) or spaced (- blurry), and can be combined with quotes (-"low quality").
• + is optional and simply marks an included term; a plain space already implies it.
• Quotes group multiple words into a single phrase. Without quotes, each word is matched independently.
• Field prefixes restrict a single term to one field: caption:, prompt:, name: (alias file:), or phrase:. Unprefixed terms search filename + prompt as usual, and captions are searched only when you explicitly write caption:, so they never dilute an ordinary prompt search. An unrecognised prefix (e.g. steps:30) is treated as a literal term.
• phrase: is the forgiving one. It searches the prompt like prompt: does, but each space or hyphen matches any single character, so phrase:"arabic eyeliner" also finds the Danbooru-style arabic_eyeliner. This is what the phrase panel uses. Use prompt: when you want an exact match.
• Terms are otherwise matched literally, so wildcard characters (%, _) and hyphenated words like close-up are treated as-is.

Viewing an image

Click any thumbnail to open the lightbox. The right panel shows:

• Tags (add/remove them right at the top)
• Dimensions, size, date, and folder
• Prompt, Negative prompt, and Template sections, each with its own Copy button
• A collapsible Caption section with a Generate caption button (see AI captions)
• EXIF / Metadata — the generation params come first in a fixed order (Model, Sampler, Steps, CFG scale, UNET, LoRA networks, Seed), then everything else
• A collapsed UserComment (raw) section holding the unparsed metadata string
• Star rating control (top bar)

The lightbox top bar also has a 🔍 Similar button — see Find similar images.

Reading long prompts

The metadata panel is built for triaging prompts:

• Collapse / expand any of Prompt, Negative prompt, Template, and UserComment by clicking its heading. The collapsed/expanded state is remembered across images and sessions.
• Resize an expanded prompt field by dragging its bottom edge, and resize the whole panel's width by dragging the divider between the image and the sidebar (also remembered).
• LoRA tags (<lora:…>) are highlighted in light green and wildcards (__word__) in cyan, in all three prompt fields.
• An empty or missing Negative prompt is hidden entirely, and the Template is shown as its own section rather than being mixed into the negative prompt.

Navigate with the ← → arrow buttons or keyboard arrow keys. Navigation spans the entire current result set — additional pages load automatically as you move past the images already fetched, so you can arrow all the way to the last (or first) image without scrolling the grid first. The counter shows your position within the full total (e.g. 198 / 79158).

Lightbox keyboard shortcuts

Key	Action
← / →	Previous / next image (loads more as needed)
Space	Advance to the next image
0–5	Set the star rating (0 clears it)
F	Toggle 5★ on/off
I	Show/hide the metadata panel
P	Pause / resume the slideshow
Del	Delete the current image (to trash) and jump to the next
Esc	Close the lightbox (also stops a slideshow)

Each prompt section (Prompt, Negative prompt, Template) has its own Copy button, and every tile shows a copy-prompt button (⧉) on hover — one click copies that image's positive prompt to the clipboard.

Slideshow

Set the two controls in the toolbar, then press ▶ Slideshow:

• Scope — This view plays exactly what the grid is showing, so it respects the selected folder, search, star filter and facets. All images ignores every filter and plays the whole library. Selected plays only the images you've ticked, in grid order (greyed out until you select something).
• Speed slider — 1 to 20 seconds per image.

Playback opens the lightbox and starts at the image you already have open, or the first one you selected, or the very first image. When it reaches the end it loops back to the start.

While it runs: ⏸ Pause / ■ Stop sit in the lightbox top bar, P or Space pauses and resumes, the arrow keys still step manually, and Esc closes the lightbox and stops the slideshow. Rating with 0–5 keeps working as images go by.

Press I first to hide the metadata panel (and B before opening, to hide the folder tree) for a clean full-screen show.

Themes

The Theme picker at the right of the header switches the visual style. Your choice is remembered, and applied before first paint so a reload doesn't flash the other theme.

Theme	Look
Classic	The original dark UI — compact, 6px radii, solid panels
Modern	Hypermodern dark: near-black #080808, hairline translucent borders instead of shadows, radial backlighting, fluid type, and lift-on-hover micro-interactions

Themes are plain CSS. client/style.css holds the base and Classic; client/theme-modern.css layers Modern on top, scoped entirely to html[data-theme="modern"], so it has no effect until selected. To add your own, copy the modern file, scope it to a new data-theme value, and add an <option> to the picker in App.jsx.

The Modern theme honours prefers-reduced-motion — the lift and scale effects are dropped if your system asks for less animation.

Hiding the panels

Both side panels get out of the way when you want the screen for the images:

• Folder tree — the ‹ Hide button at its top, or the B key. A thin strip stays on the left; click it to bring the tree back.
• Metadata panel — the ℹ Info button in the lightbox top bar, or the I key.

Both remember their state across sessions.

Where you left off

The grid remembers its scroll position and restores it when you come back — per view, so changing folder, search or filters starts at the top as you'd expect, while returning to a view you were reading puts you back in place. If the saved spot is deep into the library, imgmgr pulls the pages needed to reach it (bounded, so a stale position can't page through everything).

Prompt phrases

At the bottom of the folder sidebar, Prompt phrases lists the phrases that recur most often across your library, with the number of images using each. Click one to filter the grid by it. It's a way to browse what your prompts actually contain instead of guessing at a search term — and to notice concepts you'd forgotten you use.

• Filter phrases… narrows the list by substring.
• 2+ words only (on by default) hides bare nouns like hair or eyes in favour of phrases that describe something: regency era, blurred background, arabic eyeliner.
• Rebuild re-reads every prompt. Do this after importing a batch of new images; the index is not updated automatically by a scan.

Clicking a phrase writes nothing — it only fills the search bar with a phrase: term, which you can then edit or combine with anything else.

Turning a phrase into a real tag

Hover a phrase and click + to promote it: the phrase becomes an ordinary tag, applied to every image whose prompt uses it. It then appears in the tag dropdown and on each image's tag list, and can be filtered like any other tag. The + turns into a ✓; click that to remove the tag again from every image.

This is deliberately one phrase at a time. Auto-promoting everything would mean 46,977 tags across 6.7 million rows on a library this size — the tag dropdown would become unusable and the handful of tags you chose by hand would be lost in it. Tags are worth most for what the prompt cannot tell you (which person, which shoot, which keeper); the phrase panel already covers what the prompt does say.

Promoted tags are not renamed or prefixed. A tag is a tag — regency era searches, filters and displays exactly like ballroom. imgmgr only records where each tag row came from, in a source column, so that:

• removing a promoted tag takes out the whole batch in one click, and
• if you had already made a tag by hand with the same name, your version survives that removal.

Promoting a common phrase can write tens of thousands of rows in one click (blurred background = 14,671 images). It is instant and fully undoable, but it is a real change to your library, not a filter — so above 500 images imgmgr asks you to confirm first. Smaller promotions go through immediately.

How phrases are found. Prompts come in two dialects, often in the same library: tag lists (1woman, blonde hair, forest) and prose (…a woman with long red hair in a messy high ponytail…). Both are cut on commas first, since grammar commas chunk prose much as tag commas chunk a tag list. A short chunk is kept whole, because it is already a concept. A long one is split at function words — a, with, in — into runs of content words, and each run contributes itself plus its endings:

"a woman with long red hair in a messy high ponytail"
  -> woman | long red hair | red hair | hair | messy high ponytail | high ponytail | ponytail

Endings and not every fragment, because in English the noun comes last: red hair is a concept, long red is not. No grammar rules and no language library are involved beyond that — frequency decides. A real concept recurs thousands of times across the library; an accidental word pair does not, and a phrase seen fewer than three times is dropped. Finally, a phrase that is only a fragment of an equally common longer one is hidden: if regency era appears 32,735 times and era 33,003, you see just regency era.

The counts are a floor, not an exact figure — clicking a phrase can return somewhat more images than the number shown, because the search also matches the phrase inside longer ones. It will not return fewer.

Building the index takes about 4 seconds per 100,000 prompts and happens automatically a few seconds after the server starts, but only the first time (or after an upgrade changes how phrases are cut).

Find similar images

Open any image in the lightbox and click 🔍 Similar in the top bar. imgmgr compares the image's perceptual hash against the whole library and shows every visually similar image, nearest first. A banner indicates you're in similar-search mode — click ← Back to all images to return.

The hash is a DCT-based pHash (a 2-D discrete cosine transform of the image, thresholded against its median) that keys on overall composition rather than fine pixel detail — well suited to spotting different seeds of the same prompt. The banner also has a Strict ↔ Loose slider that adjusts the match threshold: drag toward Strict for near-identical images only, or toward Loose to include looser compositional matches. Your setting is remembered across sessions.

AI captions

imgmgr can generate a natural-language caption for an image by sending it to an SDNext server's vision endpoint (/sdapi/v1/vqa). Point sdnextUrl at your running SDNext instance and pick a captionModel in config.json (see the config table); the model must be available on that server.

• Single image — open the lightbox and expand the Caption section in the sidebar, then click ✦ Generate caption. The result is saved and shown there (collapsible, with its own Copy button); use ↻ Regenerate caption to replace it.
• Many images — select images in the grid and click ✦ Caption N in the toolbar. They're captioned one at a time with a live Captioning 3/12… progress indicator; each caption is stored as it completes.

Captions are stored in the database and can be searched with the caption: prefix in the search bar (see Search syntax) — for example caption:"stone bridge".

Note: captioning is a heavy operation on the SDNext side. The first request after the model loads can take a minute or so; subsequent ones are faster. imgmgr downscales each image before sending it to keep the request small.

Rating images

Click the stars in the lightbox sidebar, hover a tile and use the star overlay directly on the grid, or press the 0–5 keys (or F for 5★) while viewing an image. To rate many at once, select images and use the Rate: star control in the toolbar (the 0★ button clears the rating on the selection).

Multi-select and bulk actions

• Click the checkbox button in the top-left corner of a tile to select it (or click again to deselect).
• Use All / None in the toolbar to select or clear the whole page.
• With images selected, Rate: sets a star rating on all of them, Move N → opens a folder picker, Tag N → adds/removes a tag, ✦ Caption N generates captions for the whole selection (see AI captions), and Delete N moves them to the trash.

Starred images are protected from deletion. Any image with a rating of 1★ or higher is refused by the delete routes — in bulk delete, the lightbox Del shortcut, and the duplicate finder alike. You'll see a notice reporting how many were skipped; remove the stars first if you really want to delete them.

Trash & undo

Deletes are recoverable. Instead of erasing files, imgmgr moves them into a trash folder (under cacheDir, so it works even when the image root is read-only) and hides them from all views.

• Right after deleting, an Undo toast appears — click it to put the images straight back.
• Click 🗑 Trash in the header to browse deleted images. There you can Restore selected images to their original folders, Delete permanently (removes them from disk for good), or Empty trash.
• The duplicate finder also deletes to the trash, so its removals are recoverable too.

Moving images

Drag a tile and drop it onto a folder in the tree, or use the Move N → button in the toolbar after selecting images. This renames the file on disk and updates its database row in place, so ratings, tags, captions and metadata all follow it.

Reorganising folders outside imgmgr

You can also move or rename folders directly on disk (with mv, a file manager, etc.). Thumbnails are cached by content hash (.cache/thumbs/<md5>.webp), so they're reused no matter where a file ends up — nothing is regenerated.

The database, however, identifies an image by its path. To stop an external move from looking like "one file deleted, one unrelated file added" — which would drop the old row and its star rating, tags and caption — the scanner reconciles moves before it indexes or prunes: any file that vanished is matched by content hash against the newly-appeared files, and on a match the existing row is simply repointed at the new path. Same row, so ratings, tags, captions and metadata survive a folder move or rename.

Just run a scan afterwards (restart the server, or POST /api/scan) so the new locations are picked up; the console reports how many files were relocated. Until then the moved images show broken thumbnails, since their stored paths are stale.

Matching is by file content, so renamed files are recognised too. When several byte-identical copies move at once, imgmgr pairs them up deterministically — which copy inherits which row is arbitrary, but the content is the same either way. Scans where nothing has moved skip this step entirely, so there's no added cost.

Offline media (removable drives, network shares)

Images that live on an external drive or an NFS/SMB share don't have to be connected for you to use imgmgr. When a scan finds a file gone, its entry is flagged as offline, never deleted:

• The image stays in the grid and keeps its cached thumbnail, so you can browse, search, filter, rate and tag it exactly as before. It's dimmed and badged ⚠ offline.
• Ratings, tags, captions and metadata are all preserved. Opening one in the lightbox shows the cached thumbnail with a note explaining the original isn't reachable and when it went missing.
• A toolbar banner reports how many files are offline, with Show offline only to review them.
• Reconnect the drive and re-scan and the flag clears itself automatically — everything is exactly as you left it. If the files come back at a different path, the content-hash matching above repoints them, so a move while disconnected is fine too.

Nothing is ever removed automatically. If you genuinely want to forget images whose files are gone for good, click Delete orphaned data in that banner — it asks for confirmation, then drops those database entries and their cached thumbnails. That's the only thing that removes a vanished image.

This also means a share that fails to mount can't cost you anything: worst case every image is flagged offline until it's reachable again.

Duplicate finder

Click Duplicates in the toolbar to open the duplicate panel.

Mode	How it works
Exact	Groups images with identical MD5 file hashes
Seed	Groups images sharing the same generation seed (last number in the filename, e.g. 00042-1827738702.png)

A third Visual mode (perceptual-hash grouping) was removed. It compared every image against every other one, which meant hours of a frozen server on a large library — and the groupings weren't good. Use 🔍 Similar in the lightbox instead: it compares one image against the library and returns instantly.

Selection spans every group at once. Use the action bar at the top of the panel to Keep oldest in all, select rest (across all groups), Clear, or Delete N selected — so you can clean up every folder's duplicates in one pass instead of group by group. Each group also has its own Keep oldest, select rest button, and you can click individual tiles to fine-tune the selection. Deletions go to the trash, so they're recoverable.

NOTE: The seed duplicate detection only works automatically when you use software like SDNext with these settings:

Settings/Image Paths:

  Images filename pattern: [job_timestamp]-[seed]

Live updates

Set "watchForChanges": true in config.json and restart the server. imgmgr watches today's dated subfolder (YYYY/YYYY-MM-DD/) for new files. When a new image is detected a banner appears at the top of the grid — click it to refresh.

NOTE: this feature expects you to have a folder structure like this:

2024/
2025/
2026/
├── 2026-03-30/
├── 2026-04-23/
├── 2026-04-25/

Create a folder each day with the full timestamp, like 2026-04-25. If you use a software like SDNext to create images, you can establish this automatically with these settings:

Settings/Image Paths:

  [ ] Numbered filenames
  [x] Save images to a subdirectory
  Directory name pattern: [date]

  Base images folder: /some path here/2026 (adjust this to the current year)

Triggering a manual scan

POST /api/scan re-indexes the entire imageRoot. This is also called automatically on startup when scanOnStart is true. Scan progress is logged to the server console.

API

imgmgr is fully driven by a small JSON HTTP API (the web client is just a consumer of it). Every endpoint — listing/filtering, ratings, tags, trash, facets, similarity search, duplicates, scanning, and the SSE live-update stream — is documented in docs/API.md.

File structure

imgmgr/
├── server/
│   ├── index.js          # Express app, Vite middleware, SSE endpoint
│   ├── config.js         # Loads config.json, derives TRASH_DIR
│   ├── db.js             # SQLite schema, migrations, connection
│   ├── migrate.js        # One-time backfills (gen-param facets, pHash recompute)
│   ├── scanner.js        # Full directory scan + file indexing
│   ├── meta.js           # PNG tEXt / WebP+JPEG EXIF + gen-param extraction
│   ├── thumbnails.js     # Sharp thumbnail generation, DCT pHash, file hash
│   ├── caption.js        # AI captioning via the SDNext VQA endpoint
│   ├── duplicates.js     # Exact / seed duplicate detection
│   ├── phrases.js        # Mines recurring phrases out of prompts
│   ├── trash.js          # Soft-delete: trash / restore / purge
│   ├── watcher.js        # Chokidar file watcher (today's folder only)
│   ├── events.js         # SSE broadcast utility
│   └── routes/
│       ├── images.js     # Image list/filter, ratings, tags, trash, facets, similar, caption
│       ├── folders.js    # Folder listing and creation
│       ├── duplicates.js # Duplicate find and delete endpoints
│       ├── phrases.js    # Top-phrase list and index rebuild
│       └── tags.js       # Global tag listing and bulk add/remove
├── client/
│   ├── src/
│   │   ├── App.jsx       # Root component, state, SSE hook
│   │   ├── api.js        # fetch wrappers for all API endpoints
│   │   ├── usePersistentState.js  # localStorage-backed useState hook
│   │   └── components/
│   │       ├── Toolbar.jsx
│   │       ├── FolderTree.jsx
│   │       ├── PhrasePanel.jsx
│   │       ├── TileGrid.jsx
│   │       ├── ImageViewer.jsx
│   │       ├── StarRating.jsx
│   │       └── DuplicatePanel.jsx
│   └── style.css
├── docs/
│   └── API.md            # HTTP API reference
├── config.json
└── package.json

License

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.