Image Library por Limniverous

Ephemeral images disappear fast, deleted tweets, expired forum threads, discord messages that vanish into scroll-back oblivion. If you didn't save it the moment you saw it, it's gone.

Image Library is a Firefox extension that fixes that. Right-click any image on the web and it's saved permanently to your computer — hashed, deduplicated, and dropped into a searchable gallery that lives in your sidebar. No more digging through a messy Downloads folder hoping you remember the filename.

Search by tag, filename, folder, or even the site an image came from. Filter by the date you added it. Everything updates as you type.

Don't want to tag things by hand? Auto-tag does it for you, an on-device AI model guesses what's in the picture, reads its dominant color, brightness, contrast, and shape, and adds it all as searchable tags. Nothing leaves your machine to do it.

Already have a folder of images you want in the library? Import them in bulk. Need to get an image back out? Copy it straight to your clipboard, or reveal the file in your OS file manager, ready to paste or drag wherever you need it.

It's local-first, private, and built for the way people actually collect images from the internet: fast, informal, and a little chaotic. Image Library just makes sure that chaos is searchable.

===============

Image Library — Getting Started

Save images from the web before they vanish, and actually find them again later.

The basics

Save something: right-click any image on a webpage → Add to Image Library. That's it — it's saved, hashed (so you'll never save the same image twice), and dropped into your gallery.

Want to tag it as you save it? Use Add to Image Library with tags… instead — a small popup lets you add tags before it saves.

Open your gallery: click the toolbar icon, or View → Sidebar → Image Library in Firefox's menu.

Finding things again

• Type in the search box — it matches tags (both auto-generated and your own), filenames, folders, and the site an image came from (try typing a site name like "twitter" or "4chan").
• Click Filters for a date range — only show images added between two dates — or tick "Only images with no tags of your own" to see what still needs a human pass, regardless of what auto-tagging already gave it.
• Everything updates as you type, no need to hit enter.

Organizing

• Select images by holding Ctrl (or Cmd on Mac) and clicking them, or click the Select button first if you'd rather not hold a key down. Shift-click selects a range.
• With images selected, Clear auto or Clear tags wipes auto-tags or your own tags across the whole selection at once — each independent of the other.
• Hover any image for a small toolbar:
• ✎ — edit its tags by hand
• ✨ — auto-tag it (see below)
• ⧉ — copy it to your clipboard, ready to paste
• ⤢ — reveal the actual file in your file manager

Auto and manual tags are separate

Each image has two independent sets of tags: auto tags (whatever ✨ generated) and your tags (whatever you typed). Search looks at both together, but you can edit or completely clear either one without touching the other — click ✎ to open a small editor with both fields side by side, each with its own Clear button. Re-running ✨ replaces the auto tags with a fresh result; it never touches your own.

Auto-tagging ✨

Click the sparkle on any image and it'll guess a few tags for you. Two things always happen: a "what's this a picture of" guess, and a set of quick reads — dominant color, whether it's bright/dark or high/low contrast, and its rough shape (square, portrait, landscape, etc). Select multiple images and use the Auto-tag button in the selection bar to do a whole batch at once — a Stop button shows up while it's running if you want to cancel partway through (it'll finish whatever image it's currently on, then stop before starting the next). Batches automatically skip any image that already has 10 or more auto-tags, so re-running it over a mixed selection only does the work that's actually needed.

Where the "what's this a picture of" guess comes from is up to you — click ⚙ to choose (opens as a full-screen settings view over the gallery; hit Done to get back):

• Built-in (default) — an on-device model, nothing ever sent anywhere. Downloads once per session (~13MB) the first time you use it, so give it a moment. It's general-purpose, not trained on memes or screenshots specifically — great at "cat," "car," "sunset," less useful on abstract or edited images.
• Local Ollama — point it at a vision model you're already running locally (e.g. one of the Gemma 4 variants) for richer, more specific descriptions than the built-in model gives. Needs a one-time setup step — see below.

Whichever you pick, the ✎ button is right there to fix anything it gets wrong — treat auto-tags as a starting point, not gospel.

Using a local Ollama model

Ollama refuses all requests from browser extensions by default — this isn't specific to this extension, it's how Ollama behaves for anyone. You need to set an environment variable called OLLAMA_ORIGINS to the value moz-extension://* (exactly that — nothing else in the value), then restart Ollama so it picks it up.

Windows, if you installed the Ollama app (most common):
1. Press the Windows key, search "Edit environment variables for your account", open it.
2. Under "User variables," click New….
3. Variable name: OLLAMA_ORIGINS — Variable value: moz-extension://*.
4. OK your way out, then quit Ollama from the system tray and reopen it (or restart your PC, which also works but is slower than it needs to be).

macOS, if Ollama is the menu-bar app:

launchctl setenv OLLAMA_ORIGINS "moz-extension://*"


then quit and reopen Ollama from the menu bar.

Linux / running ollama serve yourself in a terminal:

OLLAMA_ORIGINS=moz-extension://* ollama serve


Here the environment variable and the command genuinely are on one line — this is a shell-only shorthand, not something to paste into a GUI's value field.

Then in ⚙ → Local Ollama, set the endpoint (http://localhost:11434 by default) and the model name (e.g. gemma4:12b — needs to be a vision-capable model you've already pulled). Hit Test connection — it checks two things separately: whether Ollama will list its installed models, and whether it will actually run one. Listing can succeed even when running is still blocked (Ollama treats running a model as higher-risk and checks it more strictly), so a real "ready to auto-tag" result needs both to pass, not just the first one.

Tick "Unload model when done" if you'd rather Ollama free up the memory it used once you're finished tagging, instead of leaving the model loaded until it times out on its own. It only unloads once — after a single image or a whole batch finishes, not after every image in a batch — since reloading a multi-gigabyte model between every image would be far slower than just leaving it warm until you're actually done.

If it's still failing after all that, double-check the environment variable's value has nothing extra in it — no trailing text, no quotes unless your OS needs them, just moz-extension://*.

Bringing in images you already have

Click Import to pick images already sitting on your computer and add them to the library the same way — hashed, deduplicated, searchable. One thing this can't do: preserve where the file originally lived on your disk (browsers don't allow extensions to see that). What it does instead: files are dated and sorted by their own last-modified date rather than the moment you imported them — so importing a folder of old screenshots slots them into your timeline where they actually belong, not all bunched up as "just added."

Backing up your tags

⚙ → Library data lets you export and import your tags as a small file — not the images themselves, just the tagging work. Handy before a profile wipe, a reinstall, or moving to another machine.

• Export tags… saves a JSON file wherever you choose.
• Import tags… restores them. Anything currently in your library gets its tags added back right away (merged with whatever's there now, nothing gets overwritten). Anything not currently in your library gets remembered quietly in the background — so if you capture or import that same image later, even much later, its old tags reappear automatically. You don't need to re-import the same file every time; once it's been through an export/import, that image's tags are remembered for good.

Getting images back out

• ⧉ Copy is the easiest way — click it, then paste straight into Discord, an image editor, wherever. It's capped at 4MB; if an image is bigger than that, it'll automatically shrink it just enough to fit rather than fail on you.
• ⤢ Reveal opens the actual file on disk in your file manager, if you want the original rather than a copy.
• Dragging a thumbnail straight out of the sidebar onto your desktop doesn't work — that's a genuine Firefox limitation, not a bug we can fix on our end. Use Copy or Reveal instead.

Cleaning up

Select some images and hit Remove. Fair warning: this permanently deletes the files, not just the gallery entries — there's no undo, and no recycle bin/trash involved (Firefox doesn't give extensions that ability). Double-check your selection before confirming.

Where do the files actually live?

Downloads/ImageLibrary/YYYY-MM/ on your computer — automatically sorted by month rather than a folder name you choose. For images you capture from the web, that's the month you added them. For images you import, it's the month the file itself was last modified (the closest thing to "created" a browser will tell an extension — if you'd edited or re-saved the file since it was originally made, it'll land in that later month instead). If you want your own categories instead of dates, that's what tags are for now — search handles that better than folders did anyway.

Quick foibles list

• No drag-and-drop export (use ⧉ or ⤢ instead)
• Remove = permanent delete, no trash
• Auto-tag needs an internet connection the first time you use it each session (for the built-in model's one-time download) or whenever Ollama needs to reach its own model
• Auto-tag's color/brightness/contrast reads are quick approximations, not exact
• Imported files are dated by "last modified," not true "date created" — Firefox doesn't expose the real thing
• Clipboard copies are always PNG and capped at 4MB, even if the original was a smaller JPEG
• Bulk auto-tag's Stop button finishes the current image before stopping — it can't interrupt one mid-request
• Tag import always merges (adds tags back) rather than overwriting — if you'd deliberately deleted a tag since exporting, importing that same export will bring it back
• Tags survive a delete-and-recapture automatically, even without ever using export/import — the extension quietly remembers a tagged image's tags by its content, not just its gallery entry
• Ollama replies are held to a strict format now — if a model ignores the instructions and replies without the expected [tag1,tag2,...] brackets, that shows up as a visible error rather than getting silently accepted in some mangled form