Paint Cam
A real neural network repaints your webcam into a living painting — right here, on your GPU. Pick a style, point your camera, and watch yourself turn into brushstrokes at video rate. Or drop in a photo. Nothing is uploaded; the whole model runs in your browser.
◆ 100% on-device · your camera never leaves this pageThis browser can't run the painter
Paint Cam needs a modern browser. Try the latest Chrome or Edge on a desktop.
Wait — this is running in my browser?
Yes. There's no server doing the painting. Paint Cam downloads a small fast-neural-style network — the same family of models that kicked off the whole "turn a photo into a Van Gogh" trend — and runs it with ONNX Runtime Web on your GPU through WebGPU. Each style is its own tiny neural network, roughly 6.7 MB, trained to repaint any image in that style.
How a frame becomes a painting
Every frame, the current camera image is shrunk to 224×224, unpacked into a raw tensor of red, green and blue planes, and fed through the network on your graphics card. The network hands back three colour planes of the same size, which get packed into pixels and scaled up to fill the canvas — that upscaling is where the soft, brushy quality comes from. When you drag style strength below 100%, the styled pixels are blended with the original frame, so you can dial the effect from a light filter to a full repaint.
Live, or on a photo
Point your camera and it paints continuously; the frame rate depends on your GPU, so a discrete card feels smooth while integrated graphics repaint a few times a second. Prefer a still? Hit Photo (or just drop an image onto the canvas) and it styles that instead. Grab a snapshot as a PNG, or record the live canvas to a video clip. Everything you make is generated locally and stays on your machine.
How can a neural network repaint my webcam in real time in a browser?
Paint Cam loads a small fast-neural-style model (~6.7 MB) and runs it with ONNX Runtime Web on your GPU via WebGPU. Every frame is turned into a tensor, pushed through the network on your graphics card, and painted back to a canvas. No server, no queue — the model executes on your own machine.
Is my camera or photo uploaded anywhere?
No. Your video never leaves your device. Frames are read locally, styled on your own GPU, and drawn straight back to the page. The only thing fetched from the network is the model file itself, once, from a public CDN — after that it's cached.
What are the five styles?
Candy, Mosaic, Rain Princess, Udnie and Pointilism — the classic fast-neural-style models, each trained to imitate a different painting. Switching styles loads that style's network; the first switch downloads it, then it's instant.
Why does it look soft or run at a few frames per second?
The model paints at 224×224 and the result is scaled up, which gives the brushy look. Frame rate depends on your GPU. If WebGPU isn't available it falls back to WebAssembly, which is slower — the Photo mode works better there.
Do I need to install anything or sign up?
No install, no account, no API key. It's a single web page, and a sample image styles itself on load so you can see it work before turning on your camera.
I build tools like this every day.
Senior full-stack engineer, available for senior or contract work, fully remote. See the rest of the lab or get in touch.