Feyard

Our standards

Last updated: October 2026

Who builds Feyard

Feyard is an independent, single-developer project run by Weiqing Li. There is no venture funding and no ad network on the drawing board that would change how your files are handled. Every tool is designed, built, and maintained by one person who reads every support email — that is also why replies can take a day or two.

Independence is a feature here: because we do not sell your data and do not run the processing on our own servers, there is nothing to monetize about your files. The business model is a simple optional subscription for people who process a lot, not surveillance.

How the tools actually work

All Feyard tools run entirely in your browser using on-device AI. The model files (open-source weights, compiled to run on WebGPU or WebAssembly) are downloaded once and cached; your images, video, audio, and documents are processed on your own device and never uploaded. This is not a marketing claim — it is an architectural consequence: there is simply no server endpoint that receives your files.

The practical effect is that the free and paid tiers share the exact same privacy guarantee. What the paid tier changes is throughput and convenience — larger files, longer clips, bigger batches, and higher-resolution exports — never whether your data leaves the device.

The models we use

We do not train proprietary models. We integrate well-known open-source models and run them locally so you get the quality without the upload:

  • Real-ESRGAN (xinntao) — image and video super-resolution.
  • HT-Demucs (Meta Research) — stem separation for audio.
  • Whisper (OpenAI) — speech-to-text, running locally in three size tiers.
  • DDColor and GFPGAN — photo colorization and face restoration.
  • SlimSAM and LaMa — object removal and inpainting.
  • Native WebCodecs / WebGPU pipelines — video transcode, trim, crop, stabilize, and export.

Each tool page links to the specific model and its license where it matters, so you can verify provenance yourself.

How we test

Because processing happens on real consumer devices, we test against a spread of hardware and browsers — not just a reference laptop. For each tool we check: a small and a large input, a slow device (mobile, WebAssembly fallback) and a fast one (WebGPU), and the boundaries where the free tier hands off to Pro. Limits published on the pricing page are the same numbers the code enforces; we do not write them twice.

Editorial standards for our guides

Guides and comparisons are written by the developer who built the tool, then checked for accuracy against the current build. Where we compare against another product, we describe its documented free-tier behaviour and date the comparison, and we link to the vendor so you can confirm. We flag our own biases plainly: we obviously prefer the tool that keeps your files on your device.

Accuracy and limits

On-device AI is fast and private, but it is not magic. Super-resolution sharpens and infers detail rather than recovering lost information; transcription accuracy depends on audio quality and language; stem separation is good but not perfect on dense mixes. We state these limits on the relevant tool and guide pages rather than overselling them.

Contact