Most AI image and video tools work the same way: you upload a file, their server runs a model, and a result comes back. Every edit leaves your device, and you are trusting someone else's infrastructure with it. Vofy takes the opposite approach: the model runs inside your browser, so the file never leaves your machine.
That is not a marketing claim, it is an architectural one. The whole pipeline - background removal, object erasing, style transfer, enhancement, video generation, trimming, compression - executes locally through WebAssembly and WebGPU.
What changed
Running vision models in a browser was impractical a few years ago. A single segmentation pass meant seconds of blocking work on a decent laptop, and the bundle was too large to ship.
Three things moved:
- WebGPU shipped. Compute shaders in the browser mean the same kernels that ran on a CUDA box can run on the user's GPU, without a plugin.
- Quantized models got good enough. Int8 and fp16 weights cut size by 3-4x with little visible quality loss for editing tasks. A segmentation or matting model that used to be hundreds of MB now fits in a few.
- WASM SIMD and threads landed. CPU fallback is no longer a punchline when a device has no usable GPU.
The result is a class of tools that behave like desktop software while living on the web.
What changes when the compute is local
No upload step, so no upload cost. There is no queue, no progress bar, no size limit imposed by a server budget. A 40 MB photo does not stall on a slow connection; throughput is your own hardware.
Privacy stops being a policy and becomes a property. Not "we promise to delete your file" but "the file was never transmitted". For anyone editing documents, medical images, or unreleased product shots, that difference is the whole product. You can disconnect from the network and keep working.
Marginal cost per edit drops to zero. Server-side inference has a real per-image cost, which is why so many tools gate the useful features behind a subscription. When the computation runs on the user's machine, that cost disappears and the product can stay free without a catch.
The tradeoffs, honestly
Local inference is not free. First load ships model weights, so cold start matters more than API latency; caching is not optional. Device variance is real - a 2019 laptop GPU and an M-series chip are not the same experience, so feature detection and graceful CPU fallback are load-bearing. And you give up the ability to swap in a bigger model server-side whenever a better one appears.
For heavy generation the cloud still wins. For the everyday editing work most people actually do, the browser is now a perfectly good place to run the model - and a much better place to keep the file.
Try it
Vofy ships 270+ tools (filters, portrait retouching, background replacement, watermark removal, upscaling, video utilities) with no account required: https://www.vofy.art/
If you are building something similar, the interesting engineering surface is not the model, it is the load strategy: what you ship on first paint, what you lazy-load per tool, and how you degrade when the GPU is not there.
Top comments (0)