Beauti is an OBS Studio effect filter for real-time skin retouching and finishing. It finds the largest face in a source and follows it, runs small neural networks locally (a face detector, a retouching U-Net, a skin segmentation model, and face landmarks that follow the face and keep eyes, brows, and lips out of the skin mask), and composes the result on the GPU in six independent stages: Retouch, Skin Tone, Local Color Smoothing, Micro Dodge & Burn, Smooth, and Eyes. Eyes adds a small iris model. Everything runs on your machine; nothing leaves it.
| Original | Beauti, preset High |
|---|---|
 |
 |
 |
 |
Full frames
| Original | Beauti, preset High |
|---|---|
 |
 |
 |
 |
The demo images are AI-generated (Google Gemini) and show no real people. The processed versions come unedited from the plugin in OBS (preset High, High Quality, Apple Acceleration); see demo images.
Beauti 0.11.1 is for macOS on Apple silicon and Windows x64. See the changelog.
| Platform | Hardware | Software |
|---|---|---|
| macOS | Apple silicon (M1 or later) | macOS 27 or later, OBS Studio 33.0.0 or later |
| Windows | x64; for NVIDIA acceleration a supported NVIDIA RTX GPU (Ampere or newer) with a CUDA 13-capable driver and NVIDIA's runtime files, otherwise the CPU | Windows x64 (validated on Windows 11), OBS Studio 33.0.0 or later |
Windows also requires the current x64 Microsoft Visual C++ Redistributable.
Intel Macs are not supported. One Windows package serves every machine. NVIDIA
acceleration needs four files from NVIDIA, which you add yourself; without them new
filters run on the CPU, which is considerably slower. OBS Studio 33.0.0 is the build baseline on both platforms;
current validation coverage is recorded in release readiness. The
Windows packages use the OBS 33 plugin folder layout (plugins\beauti\beauti.dll
with data\), which older OBS releases do not load.
The installation guide has the full steps, including NVIDIA acceleration on Windows.
macOS. Quit OBS, extract the package, and copy beauti.plugin into
~/Library/Application Support/obs-studio/plugins/.
Windows. Quit OBS, extract the package, and copy the beauti folder into
%ProgramData%\obs-studio\plugins\. For NVIDIA acceleration, download NVIDIA's
runtime archive and copy its four DLLs next to beauti.dll, then restart OBS.
Select a source and choose Tools → Beauti: Add Filter to Selected Source, or add Beauti under the source's effect filters.
Choose a Preset: Low, Medium, High, Very High, or Manual. New filters start at Medium. Outside Manual only the preset and the processing controls are shown. Manual shows every effect section, starting from the values of the last selected preset; your edits are saved with the scene. Selecting a preset replaces the effect values and never changes the processing settings. See the preset table.
| Control | What it does | Medium |
|---|---|---|
| Retouch | Reduces blemishes with the retouching model | 70% |
| Skin Tone | Evens out regional skin chroma | 70% |
| Local Color Smoothing | Reduces small skin color variations while retaining lightness | 30% |
| Micro Dodge & Burn | Reduces small lightness variations | 55% |
| Smooth | Reduces skin-detail contrast (Texture) | 40% |
| Eyes | Subtly brightens the iris and eye whites, preserving natural color | 50% |
Stages run in the order shown. Each control ranges from 0 to 100%; zero disables only that stage, and with all effects at zero the analysis stops after two seconds and the models release their memory after three minutes. Changes fade in over 150 ms so dragging a slider never flashes the unprocessed image. Eyes strength follows the selected preset; its iris model runs only while Eyes is enabled or fading out. Closed, too-small, or uncertain eyes receive no correction. See Eyes for behavior and limits.
High is the reference look. Very High deliberately removes more fine skin texture while keeping the same eye strength.
In Manual, Smooth exposes four detail reductions, Structure Size, and Edge Protection; Local Color Smoothing exposes area size and edge protection. Their values come from the last preset. Strength 0 disables each effect, without extra switches. See Texture and Local Color.
Further settings:
- Quality: Performance (384 px retouch) or High Quality (512 px retouch). The skin mask always uses 256 px. High Quality costs roughly twice the retouch time.
- Analysis Resolution: the longer side of the image the models analyze (960, 1280, 1920 px). The output keeps the source resolution.
- Processing Device: Apple Acceleration (Core AI: Neural Engine and GPU), NVIDIA Acceleration (optionally Full Precision), or CPU. New filters start on the accelerator when the machine has one and on the CPU otherwise, which the status line then says. If a selected device is unavailable, the status line reports the error instead of silently switching.
- Target Frame Rate: how often a new analysis is requested. Automatic follows the canvas rate as far as the computer keeps it steady, and otherwise uses an even fraction of it, up to 60 FPS.
The status line shows the analysis size, the measured analysis rate, and the delay of the displayed frame; Refresh Status updates it. It also explains when processing is blocked by a missing source, an unavailable image size, a source larger than 4096 px on a side, or an unsupported source or canvas color space. Reopening the filter settings also reads the current status. The plugin version appears below the settings.
The interface follows the OBS language setting: English, German, Spanish, Portuguese (Brazil), French, Russian, Japanese, Korean, and Chinese (Simplified and Traditional). Missing translations fall back to English.
- One face: the largest face is processed and followed; others are left untouched. Another face takes over only when it stays clearly larger.
- SDR only: sources that are not sRGB, or larger than 4096 px on a side, are passed through unchanged. HDR is not validated.
- A failure of up to 300 ms, such as a brief occlusion, is bridged. After that the filter fades back to the unprocessed image and resumes automatically when a face returns.
- Frame rate and delay: the whole source shows the frame that was analyzed. While a face is processed, its video therefore runs at the analysis rate and arrives slightly late. On an M4 Pro the delay is about one frame at 30 FPS (34 ms), and a slower device delays more. If the picture leads the sound, set a matching sync offset for the audio source in OBS.
- 4K at a 30 FPS canvas: an M4 Pro delivers 30 processed frames per second in both quality modes.
All processing is local. The plugin opens no network connections and disables
ONNX Runtime telemetry. Compiled model caches are stored under
~/Library/Caches/Beauti on macOS and %LOCALAPPDATA%\Beauti\Cache on Windows;
on macOS, Core AI also keeps its own cache for OBS under
~/Library/Caches/coreai-cache/. Deleting these caches is safe; the next start
rebuilds them.
Beauti is developed with the support of AI assistants. They help with writing and reviewing code, tests, and documentation, and with running builds and measurements. The work is guided and supervised by people: they set the direction, make the product decisions, decide which changes are accepted, and are responsible for the result. Changes go through the same automated checks (native and Python tests, sanitizers, formatting, provenance), and image quality is judged by people on real cameras.
The repository contains the complete native source, model preparation recipes,
license records, tests, and packaging tools. Generated models and SDKs live in the
ignored .cache/ directory. On macOS, cmake --workflow --preset macos builds and
tests on a clean checkout.
- Build instructions
- Source layout, formatting, and contributing
- Architecture, the finishing algorithms, and the decisions
- Tests and the output continuity contract
- Performance and image-quality acceptance
- Apple acceleration (Core AI)
- Model preparation, provenance, and dependencies
- Release process and changelog
Copyright © 2026 wdominik. Beauti is free software: you can redistribute and modify it under the terms of the GNU General Public License, version 3 or (at your option) any later version; see COPYING. It comes without any warranty. An additional permission allows combining it with NVIDIA's runtime. Third-party code, models, and libraries are recorded in licenses and the provenance register. The retouch model (ModelScope) and the MediaPipe models (face detection, skin segmentation, face and iris landmarks) are published under the Apache License 2.0.