The shippable version. You define a fixed label set — your tag vocabulary or
your moderation policy — and CLIP scores every image against it. No dataset, no fine-tuning, no
server: the whole classifier is a list of words plus a threshold you own.
…or drop / choose your own
backend –latency –labels –
Auto-tag mode uses the raw cosine similarity per tag and a threshold, so an
image can get many tags (or none) — the honest way to do multi-label. Moderation mode uses the
softmax across mutually-exclusive policy classes and takes the top one as the verdict.
