Skip to main content

Settings → Images

Theia is the part of Mnemosyne OS that remembers pictures. This page switches it on, installs what it needs, and hosts the two engines that share its Python: background removal and local image generation. The rail entry is Images; the page is titled Image memory (Theia) and its subtitle reads Your images become memories: find them again by meaning, color or lookalike — all local.

Every control on this page acts as soon as you click it. There is no Save button. The page opens with Loading status…; if the status cannot be read, Could not read the Theia status. and Retry.

For what image memory does once it runs, see Theia and Image memory; for the pictures in the Gallery and the Image studio, see their own pages. This page is about the settings.

Activation​

Enable image memory: Watched folders start ingesting images (png, jpg, webp…). Off = nothing changes. Off by default, saved at once.

Opt-in, and only for new folders

With the switch off, nothing about images changes. Switching it on adds the image extensions to the new folders you attach in Configure DocWatch; a folder attached earlier keeps its extension list until you edit it with ✎.

Setup​

Shown once image memory is on. Three steps in a ladder; a done step reads ✓ Ready, an actionable one shows a button, a running one shows a spinner, the stage (Preparing environment, Installing dependencies, Finalizing) and a percentage, with a violet progress bar under the ladder. Jobs run in the background and survive leaving the page, one at a time: Another Theia job is already running — wait for it to finish.

  1. Vision engine: Local Python runtime for the image embedder (~500 MB, one time). Button Install.
  2. SigLIP 2 model: The image model weights (~1.5 GB), pinned to an exact version. Button Download, available once the engine is ready.
  3. Index your images: Computes vectors, thumbnails and color palettes for images already in your vaults. Button Index now, available once engine and model are ready. This step never shows as done: it is a pass you can repeat.

While a pass streams, a line counts it: the vault, n image(s) indexed, done/total (pct%) when a total was measured (never a guessed percentage), and n failed when some were.

With the engine and model ready and nothing running, a status box reads one of: Index pass running…, Pass scheduled — starting shortly, Last pass with its time and n image(s) indexed (plus, when relevant, n refreshed without re-embedding, n failed, n missing file(s)), or No pass yet this session. Under it: Indexing is automatic — at startup, when images arrive, and on the first recall. This button just forces an immediate pass. The page refreshes every three seconds while a pass is scheduled or running.

Performance​

Shown when a CUDA-capable card is detected or the engine already runs on CUDA, and the engine is ready. GPU acceleration (CUDA): Embeds your images on the GPU (name): several times faster on large imports. followed by The CUDA runtime is already on this machine: no download needed. or One-time CUDA runtime download (~2.4 GB). The button is Switch to GPU. The step reads ✓ Ready only once the engine has proven it runs on CUDA.

During the switch

The runtime is shared with the voice engines, so a machine that already paid the 2.4 GB for one of them installs it offline. While the swap runs, every image job refuses to start; an indexing pass in flight stops cleanly and resumes on the GPU.

Documents​

Shown once image memory is on. Extract the images inside Word documents: Creates a _mnemosyne-images folder in each watched folder, with the pictures found in your .docx files. Off = nothing is written. Off by default: it has its own switch because it writes real files into your folders. Switched on: Documents already in memory are read at the next startup: their content has not changed, so nothing re-opens them before then.

The extracted files land in _mnemosyne-images/<document>/ inside the watched folder itself, never in an app folder, and that folder is skipped by the watch so that each picture enters memory once.

Engine​

Once image memory is on and the engine is warm, a line reads Engine: followed by the device (or — when unknown) and the model id.

Background removal​

Shown as soon as the engine is ready, even with image memory off. Two cards, Standard (BiRefNet lite · 224 MB · a few seconds per image, the default) and High quality (Full BiRefNet · 973 MB · the best edges for photos), saved at once. Each card adds · downloaded at first use until its weights are on disk, then ✓ Ready.

Used by right-click → Remove background on canvas stickers and photos. The chosen model downloads at first use. Images of 768 px or less always use Standard: at that size the result is identical and much faster.

Both models are MIT-licensed and the matting runs on the CPU. The gesture itself is on the Canvas.

Generate on this machine​

Shown once the engine is ready: the local image generation engine, the same panel the Image studio folds away in its rail. It borrows the vision engine's Python.

  • Loading the models… while the list loads; Could not read the model list. and Retry if it fails.
  • A memory chip while a model sits in RAM: its size in GB, or In memory when the size cannot be read (never 0). The model is loaded in RAM so the next picture starts at once. It lets go on its own after ten idle minutes. Release frees it now.
  • A machine line: n GB of RAM, m of video memory, or an unreadable amount of video memory when the app cannot read it.
  • When the dependencies are missing: Install the local engine, then Installing with a percentage and the stage.

Hugging Face account​

Shown when the catalogue holds a gated model or a token is stored. Some models refuse to download without an account of your own. Mnemosyne keeps no copy of them: you accept the licence on the vendor's page, and lend a token you can revoke.

  • Not connected: A read token is enough. Do not paste one with write access., a Create a read token link that opens the Hugging Face token page in your browser, a password field (hf_...) and Connect (Checking…).
  • Connected: ✓ Account connected, then Check and Disconnect.
  • Verdicts: Connected as name; Hugging Face refused this token. Create a new one and connect again.; Could not reach Hugging Face, so the token is saved but untested. (an unknown, not an error); That does not look like a Hugging Face token, which normally starts with hf_. Saved anyway.
  • Refusals before anything is saved: Paste a token first., a token containing whitespace, That is a web address, not a token…, and This machine refused to store the token safely, so nothing was saved.

The token is sealed in the OS keystore and never read back; the field is cleared once it is saved.

On a Linux keyring that cannot seal

When the system keyring is the plaintext basic_text backend, a red line says so: On this machine the system keyring seals secrets with a key that is publicly known, so treat this token as stored in the clear. The page never hides this.

Model catalogue​

A sort bar: By category (default), For this machine, Fastest, Best quality. Categories, empty ones hidden: Illustration, Anime and manga, Photographic, Fast, less detail, General purpose, Top tier, for a workstation.

Each row: the name, a suggested chip on the recommended model, the size in GB, and how it fits this machine: runs fast here, runs slowly here, too big for this machine, or speed unknown. A gated model reads Needs your Hugging Face account, carries an Open this model's page on Hugging Face button, and stays disabled until an account is connected (This model is gated. Connect your Hugging Face account, then accept the licence on the model's page.).

  • Download counts pct% · stage with a Stop button: Stops the download. What has already come down is kept, so it resumes where it left off. A stopped download reads Stopped with Resume; an interrupted one shows Retry.
  • Ready: ✓ Ready and a bin. The bin takes two clicks (Delete?): Removes size from this disk. It has to be downloaded again., then n GB freed.
  • A download requested while another job runs waits: One job at a time — this starts when the running one is done. The licence of each model is the tooltip of its download button.

Your own models​

Your own models lists what you added, each with a Remove button, and a field, owner/model, or a folder on this disk, with + to add one.

The line at the bottom​

Everything runs on this machine: the model reads your images locally, and thumbnails are stored inside the vault, protected like the memories they belong to. Errors from any action print at the bottom as ⚠ and the message. On Windows, a refused engine start reads Windows refused to start the image engine. Restarting Mnemosyne clears this; if it comes back, restart your computer.

What does leave the machine on this page: a Hugging Face token, sent to Hugging Face when you connect or check it and when a gated model downloads; the model downloads themselves. Your images never do.

Traps​

  • Image memory is opt-in and only touches new watched folders; edit old ones by hand.
  • Document-image extraction writes files into your watched folders; it is its own switch, off by default, and documents already in memory are caught up at the next startup.
  • The GPU switch downloads about 2.4 GB once and is never marked done without proof that CUDA works.
  • Gated generation models need your own Hugging Face account and your own acceptance of the licence, on Hugging Face.
  • Deleting a downloaded generation model is a two-click, irreversible disk removal.