In July I wrote about teaching an app to see clothes — the Core ML recognition inside WRDB, my wardrobe app. This lab is its web sibling, built around one question: can the same idea run in a browser so that privacy isn't a promise in a policy, but an architecture? If the photo never leaves your machine, there is nothing to trust. There is no server.
The result is an interactive page where you drop a photo of a garment, it labels itself in under a second, and a live network counter — armed the moment the model finishes loading — stays at zero while you use it. That counter is the whole thesis in one UI element.
Zero-shot, and why the vocabulary is code
The model is fashionSigLIP, a CLIP-style dual tower fine-tuned for fashion, quantized to int8. There's no training anywhere in this lab: classification is zero-shot, comparing your photo against 19 frozen English prompts like "a photo of a pair of denim jeans".
Those prompts turned out to be the most sensitive dial in the system. Adding two missing classes and changing one word — "jeans" to "denim jeans" — moved pilot accuracy from 70% to 100%. Which is exactly why the vocabulary is version-locked like code: change a prompt and every published number is invalid until you re-measure. My favourite consequence: the browser never even downloads the text tower. The 19 label embeddings are precomputed by the eval harness and shipped as a 100 KB JSON, so the demo pulls only the vision tower — 94 MB, cached by the browser after the first visit.
The dataset is the interface
The eval set is 59 photos of real garments from my own wardrobe: on the bed, on a rug, wrinkled, at night. That realism is not a flourish — it is the measurement. Early on I ran a pilot on background-removed cutouts and got 91% top-1. The same model on raw photos of the same kind of clothes: 70%. Twenty-one points that separate a benchmark from the thing a user will actually upload.
The final numbers, measured in the browser (the runtime you actually use, more on that below): 85% top-1 (95% interval: 73–92) and 100% top-3 on n=59. That second number decided the product: the right label never once left the model's first three guesses, so the demo shows you three answers, not one.
The nine misses are published on the page, photos included, and they tell one coherent story. Long sleeves drift towards "sweatshirt". Chunky knits blur into fleece. And all three cargo trousers photographed folded were read as shorts — the framing fooled the model, not the garment. The dataset is the interface: how people photograph clothes matters as much as what the clothes are.
Quantized models disagree with themselves
An int8 model is not one model — it is one set of weights executed by whatever kernels the runtime provides. The same photo, same weights, gave different answers in Node and in the browser on 5 of 59 photos, and one borderline t-shirt collected three different labels across three backends.
The fix is methodological, not technical: measure in the runtime you ship. The published accuracy comes from a Playwright-driven run inside a real browser, backend declared in the metrics file, and the harness treats a dev-mode run leaking into the official numbers as an error.
The GPU lost to WASM
My favourite finding. The plan said: WebGPU for speed, WASM as the honest fallback. Then I measured on an actual GPU (Apple metal-3): ~1.6 seconds per photo on WebGPU vs ~0.7 on plain WASM. For int8 models the dequantize overhead eats the GPU's advantage, and SIMD on the CPU quietly wins. Bonus trap: headless Chromium had been reporting "WebGPU" while silently running SwiftShader — a software renderer — which is why the adapter identity now gets recorded next to every latency number.
So the demo runs WASM on purpose. The fallback won, everyone gets the fast path, and one whole class of "does your browser support X" failure modes disappeared.
What this is — and isn't
The set is small and declared: 59 photos, 16 of 19 classes covered — coats, dresses and vests simply had no photos, and those cells stay honestly empty. Ten photos came from other people's wardrobes and count in the metric, but only my own photos appear on the page. Zero-shot means nothing here was tuned to my closet; it also means a fine-tuned model would likely beat it. And the 95% interval is wide — n=59 buys you 73–92, not a decimal point.
With more time: the q4f16 weights (54 MB, needs re-measuring), a bigger set to tighten the interval, and the comparison this lab was born for — the same photos through WRDB's Core ML pipeline, on-device against in-browser.
Everything reproducible — harness, frozen vocabulary, test set and the zero-network Playwright test — ships with the lab. The page is the product; go feed it a photo of whatever is on your chair right now.