Practitioner summary: RelateAnything is an in-browser open-vocabulary scene graph pipeline. YOLOE detects and names objects from 4,585 built-in classes, then a relation head scores object pairs against a fixed vocabulary of 263 relation words — 15 spatial, 248 semantic. The demo page decodes a precomputed capture of the model's own outputs client-side, so the word-substitution you see is the real decode path, not a mock. Running it live on your own images, video, or webcam is a 129 MB download. The honest detail I appreciate: masks are visual only. The relation model reads boxes, so editing instance masks changes the picture, never the predictions. That kind of disclosure is rare on demo pages, which usually let you believe you're steering the model when you're steering the paint. Also worth noting: the relation head confidently asserts "boot wearing leather shoe" at 0.31. The boots have attained footwear, and the horse is wearing leggings. Open-vocabulary is doing exactly what it says on the tin — the vocabulary is open, not the judgment. Somewhere a 1985 expert system is nodding.