About Feyard

AI that runs on your device

Feyard is a collection of free AI tools that run as local models inside your browser. Processing happens on-device through WebGPU and WASM. No data leaves your machine.

No account, no upload, no server-side inference. Every image you upscale, audio file you transcribe, and PDF you summarize is processed on your own hardware. Files never touch a cloud server.

Why local?

• Privacy by default. Because data never leaves your browser, there is nothing to leak, breach, or sell. The tools suit sensitive documents, medical records, legal contracts, and personal photos.
• No per-use cost. Cloud AI tools charge per API call. Browser inference has none — process 10 files or 10,000 and the price is the same: nothing.
• Works offline. After the first model download, most tools keep working without an internet connection.
• Open standards. The tools build on WebAssembly, WebGPU, and open-weight models (Transformers.js, Tesseract, Kokoro, RNNoise) — no proprietary SDK and no vendor lock-in.

What we build

Feyard is twenty browser-based tools spanning image, audio, video, text, and document work — background removal, upscaling, compression, photo restoration, OCR, speech-to-text, text-to-speech, PDF summaries, document search, vocal separation, vectorization, and more.

Each is built as a full alternative to a paid cloud service — not a demo or a teaser, but a complete product that runs at native speed in your browser.

How the tools work

Depending on the tool, Feyard uses different browser technologies: WebGPU for GPU-accelerated neural inference where available, with a WebAssembly fallback everywhere else; open-weight models such as Whisper (speech-to-text), Swin2SR (super-resolution), GFPGAN (face restoration), LaMa (inpainting), DDColor (colorization), BiRefNet and MODNet (matting), Tesseract (OCR), Kokoro (text-to-speech), and RNNoise (denoising) — all under MIT or Apache licenses and attributed on the tool pages that use them.

Models are downloaded once from a CDN on first use and then cached in your browser's local storage, so later sessions work from the cache — most tools keep functioning offline. Each tool page documents the exact model it uses, its approximate download size, supported input and output formats, and the hardware it needs.

Known limitations, stated plainly

On-device processing trades convenience for hardware: large files can fail on low-memory devices, WASM fallback is slower than WebGPU, and every generative or reconstructive feature (object removal, restoration, upscaling, depth estimation) can produce artifacts — the model reconstructs missing information, it does not recover the original pixels. We document these limits on each tool page rather than overselling results, and we mark tools that are still in development as such until they are complete.

Who we are

Feyard is an independent project maintained by one developer. It takes no venture capital and has no plans to. The site is supported entirely by non-personalized advertising — ad slots are kept out of tool workspaces and pages without substantive content, and advertising cookies stay off unless you explicitly accept them.

Contact

Questions, feedback, and bug reports go to hello@feyard.com.