AI

Everything on this page runs on your own computer, and no photograph, face or name is sent anywhere by it. The one exception is an extension you switch on that says otherwise on its switch (Gemini, below). The models themselves are downloaded once, when you turn their feature on.

Two kinds of computer

Ninaivu is tested on two kinds of machine, and works out which yours is when it starts (System → Performance says which, and why):

	Basic	Full
The machine	anything else: a mini PC, an old laptop, a single-board computer, a desktop without a graphics card or without the image model	a graphics processor with at least 6 GB of memory, or Apple silicon, or any computer with 8 GB of memory once search by description is added
The gallery, albums, sharing, roles	yes	yes
Faces, grouped and named	yes	yes
Sideways photographs put right	yes	yes
Places named from the location	yes	yes
Text read from photographs (with the text reader)	yes, slowly	yes
Search by words in names, dates, places and people	yes	yes
Mugil, the encrypted copy	yes	yes
Tags, "the beach at sunset" search and descriptions from the image model	—	yes
Finding visually similar photographs (S in the viewer)	—	yes
Videos described from several moments	—	yes
Creative Studio: generative edits, object removal, upscaling	—	with a graphics card

On a Basic machine, auto means the light engine, so it never tries to load the large models and starts quickly. Choosing the image model outright (--ai clip, or ai_engine on Advanced settings) still loads it. Setting hardware_tier on Advanced settings overrides the measurement.

Adding search by description

It is not installed by default: it is about 2 GB, plus the model itself. Add it from the first-day walk-through or Home & extensions → Extras → Search by description, then restart Ninaivu. On a Full machine it is used from then on; the model's weights are fetched once at that first start, and after that Ninaivu tells the model libraries to make no request at all (HF_HUB_OFFLINE).

What each pass does

AI → AI models has a switch for each, saying which model it needs and whether that model is here.

• Faces — a 37 MB detector and matcher. Groups are named on Faces; name one and the rest follow.
• Places — an 11 MB list of places, downloaded once from GeoNames; the photograph's own location becomes "Paris" or "the beach". The map uses a built-in outline of the world; street-map tiles from OpenStreetMap are off unless map_tiles is switched on, because each tile asked for shows roughly where a photograph was taken.
• Sideways photographs — a 77 MB model that says which way up a photograph goes. By default it looks by itself after every scan and turns what it finds straight away, as one batch that Undo on Review → Straighten puts back; switch that off and what it finds waits there for approval.
• Text in photographs — the text reader from Extras; slow on a processor.
• The image model (Full, once added) — CLIP or SigLIP: tags, "the beach at sunset" search, descriptions, visually similar photographs.
• Videos (Full) — described from several moments rather than the poster frame.

Sudar

Sudar (சுடர், glow) edits in the browser: light, colour, straightening, crops, looks, and suggestions measured from the picture. "Make it warmer" in plain words. The original is never changed; a copy is saved beside it, and so can any result a tool makes — a cut-out, a colourised print — a family member's copy waiting for the administrator's approval like an upload. The Photo Studio (Edit photo in the viewer) renders through the same engine, so a slider means the same thing in both; the family guide has every panel.

Its Skin and Hair tools work on every person in the photograph, each judged against their own skin, using the face model from AI models. They leave a bindi, sindoor or sacred ash exactly as it is, and have no setting that lightens a complexion. Hair the colour of what is behind it is found with the help of Background removal (RMBG-1.4) when that model is installed, and painted in by hand otherwise; Add hair draws new hair from the strands beside it, on this machine. See the family guide.

Its Enhance tools each need a model, and the card says which:

• Upscale — Upscale ×4 (Real-ESRGAN), from AI models.
• Restore faces — Restore faces (GFPGAN 1.4), from AI models.
• Colourise — a DDColor ONNX model. The catalogue cannot pin one yet, so this is the one model placed by hand: put ddcolor.onnx in the onnx subfolder of the models folder (AI models shows the folder), or point NINAIVU_COLORIZE_MODEL at the file. With Restore faces too, Revive old photo does both in one tap.

Each runs on this machine and nothing leaves it; with the Creative Studio extension, an AI server workflow can take any of them instead.

Extensions

Anything that cannot keep the core's two promises — nothing leaves the house unless you chose where, and it runs on a small machine — is an extension, off until switched on under System → Home & extensions → Extensions (it takes effect when Ninaivu next starts), and its switch says what it does:

• Creative Studio — generative edits and the heavy tools with large models, on this computer with a GPU or on a ComfyUI server at home. The AI server page refuses an address on the internet, so nothing leaves the house (a Tailscale address counts as home).
• Gemini — edits and descriptions from Google's models. A re-encoded copy of the photograph goes to Google when, and only when, a request names it. Only an administrator can send one, unless Let family members send photographs to extensions that go outside the house is ticked under Home & extensions → Extensions.

Setting up extensions is covered in the technical administrator guide.