Search models

Search is powered by a model that runs entirely on your machine. Weights are downloaded on first use of a command that needs them, never at install.

Search models

Selected with --model. Each link goes to the model card on Hugging Face, where training data, intended use and limitations are documented by the people who built it.

Model	Download	Dimensions	Notes
google/siglip2-base-patch16-224	~1.5 GB	768	The default
google/siglip2-base-patch16-384	~1.5 GB	768	The same model at 384px: finer detail, about 4x slower to embed
google/siglip2-so400m-patch14-384	~4.5 GB	1152	Largest and slowest

All three are SigLIP 2, which understands queries in many languages, Turkish included. Searched by their own captions on 300 test photos, the default found the right photo first for 84% of English queries and 52% of Turkish ones; the SigLIP 1 model it replaced managed 81% and 37%, at the same speed.

Higher resolution and more parameters generally mean better matching on fine detail, at proportionally more time per image and more disk. Whether that helps your photos is an empirical question: see using several search models.

Any other SigLIP or SigLIP 2 model on Hugging Face can be given by id, such as the older English-only google/siglip-base-patch16-224. Two SigLIP 2 kinds do not load: the -naflex models, and the giant-opt ones.

A library whose config.toml already names a default_model keeps it. videre writes that key when it creates a library, so most libraries made before SigLIP 2 became the default still use google/siglip-base-patch16-224; videre config shows which. To move one over:

videre config set model google/siglip2-base-patch16-224
videre embed

The old vectors stay on disk until you remove them; see using several search models.

Face model

Not selectable. videre faces always uses InsightFace WePrompt/buffalo_l, about 180 MB, an SCRFD detector plus an ArcFace embedder.

Both live in the shared Hugging Face cache at ~/.cache/huggingface/hub/, overridable with HF_HOME. See what gets downloaded.

scan, dedupe, fix-dates, prune, stats, and locations without similarity search need no model at all.

Choosing one

embed, search, classify, gallery and mcp all take --model <id>, resolved as --model first, then default_model in your config, then the built-in default.

videre config set model google/siglip2-base-patch16-384   # lasting default
videre embed --model google/siglip2-base-patch16-384      # just this once
videre config                                             # show what resolves

The other models are only fetched if you actually select one: siglip2-base-patch16-384 is about 1.5 GB, and siglip2-so400m-patch14-384 about 4.5 GB.

One model never disturbs another

Each model keeps its own data, so preparing a second leaves the first untouched and switching between them invalidates nothing. Only videre embed creates a model’s data; everything else reads it.

Asking for a model you have not prepared is an error listing the ones you do have, rather than silently returning nothing. videre dedupe --html is the exception: a missing model disables its in-page similarity search with a note, rather than failing a page that works without it.

For how to actually try, compare, switch and remove one, see using several search models.

Where the data is kept

Not in the main database. Each library and model pair gets its own file:

<library>/.videre/embeddings/<owner>--<model>.db

Per library rather than one shared file per model, because videre prune cannot see another library’s contents. A shared layout would let one library’s cleanup delete data another still needs, and an embedding costs hours to rebuild. Caches shows the same tradeoff decided the other way, for thumbnails.

The library part of the path includes a hash of its canonicalised path, so two libraries both called photos.db in different folders never collide.

Expect roughly 130 MB to 190 MB per model for a 70,000 photo library. videre stats reports the actual figure per model.

Upgrading from before 0.10

Data written by 0.9.x lived in the main database. That fallback was removed in 0.11.0, so such a library now reports the same clear error as any other missing model.

Nothing is deleted: the old data sits untouched and can be dropped by hand. Rerun videre embed to rebuild in the current layout.