# Longpipe

> Open-source SDK for real-time video effects in the browser: virtual backgrounds,
> background blur, portrait matting, face touch-up, auto-reframe, and microphone
> noise removal. Runs custom-trained models as hand-written WebGPU/WebGL2 shaders
> in a fully-GPU, zero-copy pipeline — no general-purpose inference runtime.
> MIT licensed (source and model weights). npm: `longpipe`.

## Essentials

- Install: `npm install longpipe`
- Public class: `EffectsPipeline` (NOT `Pipeline`) — `import { EffectsPipeline } from 'longpipe'`
- Hand it a `MediaStream`, get a `MediaStream` back: `new EffectsPipeline(stream, options)`
- Construction is synchronous; `pipeline.stream` is available immediately and emits
  unprocessed passthrough until `pipeline.ready` resolves.
- Model weights stream from `https://cdn.longpipe.dev/models/v/0.0.5/` by default; self-hostable.
- Browsers: Chromium, Firefox, Safari (desktop + iOS). WebGPU when available, WebGL2 fallback.
- Five model presets (`xs`, `small`, `medium`, `large`, `xl`); `preset: 'auto'` benchmarks
  the device at init and an adaptive controller swaps presets at runtime.
- Backgrounds, touch-up and auto-reframe share ONE encoder pass — enabling a second
  effect does not cost a second inference. Audio denoise is a separate AudioWorklet pipeline.

## Docs

- [Longpipe docs](https://longpipe.dev/docs): Open-source video effects SDK — overview and quick links
- [Getting started](https://longpipe.dev/docs/getting-started): Install Longpipe and add your first effect
- [API reference](https://longpipe.dev/docs/api): EffectsPipeline class reference
- [Architecture](https://longpipe.dev/docs/architecture): How Longpipe works — one encoder, many heads, and a fully-GPU pipeline
- [Backgrounds](https://longpipe.dev/docs/backgrounds): Accepted background input formats
- [Touch-up](https://longpipe.dev/docs/touchup): Face touch-up (skin smoothing) — options, styles, and how it composes with backgrounds
- [Auto-reframe](https://longpipe.dev/docs/reframe): Auto-reframe — framing the subject automatically, in auto or manual mode
- [Audio denoise](https://longpipe.dev/docs/audio): Real-time speech denoising (DeepFilterNet3 / RNNoise)
- [Presets & autotune](https://longpipe.dev/docs/presets): Model size knobs and how autotune picks one
- [Model layer](https://longpipe.dev/docs/model-layer): Run the Longpipe matting model directly, without the pipeline
- [Self-hosting weights](https://longpipe.dev/docs/self-hosting): Point Longpipe at your own model weight CDN

## Full text

- [All documentation as one file](https://longpipe.dev/llms-full.txt): every page above,
  concatenated as plain markdown.

## Source

- [GitHub](https://github.com/sb2702/longpipe): SDK source, training code, model export.
- [npm](https://www.npmjs.com/package/longpipe)
- [Live demo](https://longpipe.dev/demo)