# AI models on your device

> The AI models DriftEdit runs in your browser, what each one does, how big it is, and where your data goes (nowhere).

https://driftedit.app/manual/ai-models/

DriftEdit's AI tools run inside your browser tab, on your own device, through ONNX Runtime Web: on the graphics card with WebGPU where your browser has it, and otherwise in WebAssembly on the processor. Each model downloads from DriftEdit's own site the first time you use its tool, and your browser keeps it for next time (until its storage is cleared). What the models hear and see never leaves your device.

## The models
| Tool | What it does | Model | Download |
| --- | --- | --- | --- |
| **Captions from speech** | Writes what is said, every word timed, in 99 languages | Whisper base (OpenAI, MIT) | 80 MB (113 MB on a graphics card) |
| **Split into stems** | A song in four parts: vocals, drums, bass and the rest | Spleeter 4stems (Deezer, MIT) | 157 MB |
| **Isolate the voice** | Dialogue lifted out of music and effects | MRX (MERL, MIT) | 31 MB |
| **Smooth motion** | Frames drawn between the footage's own for fluid slow motion and 60 fps exports | RIFE 4.17 lite (MIT) | 11 MB |
| **Move in depth** | How near each part of a photo is, to push into it | Depth Anything V2 Small (Apache-2.0) | 27 MB |
| **Blur the background** | A person's outline, to keep them sharp | MODNet (Apache-2.0) | 26 MB |
| **Still frame** | A subject cut out of a frame, whatever it is | BiRefNet lite (MIT) | 98 MB |
| **B-roll on the words**, **Find scenes** | Matches pictures to words, and finds a video's scenes | CLIP ViT-B/32 (OpenAI, MIT) | 156 MB |
| **Perfect quality (AI)** (Pro) | Every frame of an export redrawn sharper and cleaner | Real-ESRGAN (Tencent ARC, BSD-3-Clause) | 4.9 MB |
| **Speak the captions** | An AI voice, a woman's or a man's, in the studio's languages | Supertonic 3 (Supertone, OpenRAIL-M) | 401 MB |
| **Speak the captions** in Chinese | The same, in Chinese | Kokoro 82M (hexgrad, Apache-2.0) | 93 MB |
| **Captions from speech**: who speaks | Finds the people speaking and names each turn | pyannote segmentation 3.0 (MIT) and WeSpeaker ResNet34-LM (WeSpeaker, CC BY 4.0) | 1.5 MB and 6.7 MB |
| **Captions from speech**: sounds | Captions the sounds heard where nobody speaks: applause, laughter, music | YAMNet (Google, MIT) | 15 MB |

No model is behind **Clean up the voice** or **Enhance detail**: the first is signal processing, the second an upscaler on the graphics card. Neither needs a download.

## What happens the first time
The first press of a tool downloads its model: the work shows on the clip it works on, as a percent, and beside the timecode. Later uses start at once. On a slow connection, the larger models take a while; the download continues while you edit.

## Speed
On a computer with a real graphics card, most models run on it, through WebGPU, and are fast; the stems and the voice isolation always run on the processor. Without one, or in a browser without WebGPU, they run in WebAssembly on the processor: slower, but the same result. If a model cannot run on a graphics card (some cards and drivers refuse some models), DriftEdit runs it in WebAssembly instead, by itself.

> [!TIP] When a model will not run
> See [Troubleshooting](https://driftedit.app/manual/troubleshooting/#ai-tools): the usual causes are a download interrupted, a browser without enough memory, or an old graphics driver.