Overview
The thresholds Nosmai Studio uses to judge a filter in the Performance view, with the reasoning. They are tuned for a mid-range phone. Treat them as guidance, not hard caps: the Performance view warns, it never blocks export.
Product: Nosmai Studio · Platform: macOS, Windows · Updated: 6 October 2026 · Reading time: 5 minutes
Performance thresholds
Each resource reads Good, Warning, or High against these bands:
| Resource | Warning at | High at | Why |
|---|---|---|---|
| Largest texture edge | 2048 px | 4096 px | Large textures cost GPU memory and bandwidth on a phone |
| Texture memory (decoded) | 64 MB | 192 MB | The uploaded footprint competes with the host app for GPU memory |
| Package size | 10 MB | 30 MB | A large package is slower to download and decrypt before the filter runs |
| Total triangles | 150,000 | 500,000 | High vertex counts raise per-frame cost on a mid-range GPU |
| Render passes per frame | 26 | 40 | Each pass is a full render-target bind and draw; passes include the engine's own stages |
| Fill (kilopixels per frame) | 28,000 | 60,000 | Fill rate is the main cost on integrated and mobile GPUs |
Render passes count every framebuffer the engine binds in a frame, including its fixed pipeline (source, face mesh, each beauty stage, composite), not just your own shader passes. A plain beauty-and-overlay filter sits around 12 to 16 passes before you add anything.
Supported inputs
| Setting | Values |
|---|---|
| Preview target aspects | Source, 9:16, 16:9, 1:1, 4:5 |
| FaceMesh model | Lite (default) or Full |
| Minimum SDK version | Default 3.0.0; set per filter at export |
Supported file formats
See File formats for the import and export format tables.
Guidance
- Start beauty smoothing low (0.25 to 0.40). It is the most expensive beauty stage.
- Resize textures to the smallest edge that still looks right; a 4096 texture is a phone risk.
- Decimate dense models toward 150,000 triangles or fewer.
- Combine full-frame shader passes where you can; fewer passes means less fill.
- Check the Performance view before every export, especially on the Full FaceMesh model, which is closer to phone behaviour.