ONNX.css – Neural network inference using stylesheets

ONNX.css, a new open-source tool, compiles ONNX neural network models into CSS stylesheets, enabling browsers to run inference via the style engine. It demonstrates three models: a LeNet-style CNN for MNIST digit classification (12 ONNX nodes, 30,852 CSS properties), a VAE decoder for image generation (6 nodes, 3,170 CSS properties), and Silero VAD v6.2.1 for speech detection (48 primitive nodes, 8,188 CSS properties). The project highlights the browser's style engine as a novel inference runtime.

ONNX model → CSS stylesheet ONNX.css Choose an ONNX model, compile it into a CSS stylesheet, then let the browser style engine run inference. Compiled successfully Draw a digit. Your browser's style engine will infer the result using a neural network expressed in CSS. Model compiled to CSS Official ONNX Model Zoo / opset 8 Input canvas 28×28 tensor Draw with mouse, trackpad, or touch. CSS prediction Argmax of ten computed logits — Waiting for a stroke — — — — — — — — — — → CSS variables → getComputedStyle LeNet-style CNN 12 ONNX nodes 30,852 CSS properties 10 raw logits Generated artifact Raw compiled CSS Preview truncated Only the beginning of the stylesheet is shown. Copy or download to inspect the complete file. Model source: ONNX Model Zoo / MNIST https://github.com/onnx/models/tree/main/validated/vision/classification/mnist . Official validated mnist-8.onnx model using ONNX opset 8. Compiled successfully Explore latent space. Your browser's style engine generates images using a neural network expressed in CSS. VAE decoder compiled to CSS 6 nodes · 3 dense layers Latent position Two model inputs The controls bind directly to two CSS custom properties. CSS-generated image 784 computed pixel values 0.00, 0.00 Waiting for CSS output → CSS decoder → 784 pixels Variational Autoencoder VAE 6 ONNX nodes 3,170 CSS properties 28×28 generated image Generated artifact Raw decoder CSS Preview truncated Only the beginning of the stylesheet is shown. Copy or download to inspect the complete file. Pretrained checkpoint: pszmk/mnist-vae-latent2 https://huggingface.co/pszmk/mnist-vae-latent2 . Decoder-only fixed-shape ONNX export generated by this project. Compiled successfully Listen for speech. Load audio, choose a fixed 32 ms window, and estimate its speech probability using a neural network in CSS. Silero VAD v6.2.1 compiled to CSS Fixed 8 kHz · recurrent state Audio window 256 samples at 8 kHz Audio decoding and resampling happen in JavaScript. CSS speech estimate Silero probability output — — 0.50 Load an audio file to begin → CSS STFT + VAD → Speech probability Silero VAD v6.2.1 48 primitive nodes 8,188 CSS properties 32 ms fixed window Generated artifact Raw VAD CSS Preview truncated Only the beginning of the stylesheet is shown. Copy or download to inspect the complete file. Model source: Silero VAD v6.2.1 https://github.com/snakers4/silero-vad/releases/tag/v6.2.1 . With minor model surgery to remove dynamic shapes: fixed 8 kHz branch with prior-audio context and the LSTM lowered to primitive operators.