New AI 3D depth models + widescreen fixes
AI 3D Depth: new model lineup
The single MiDaS depth option is now a 7-way lineup you can cycle through, ordered fastest to slowest, to compare speed vs. depth quality on your own setup: MiDaS-192 (default) → YOLO26-N-256 → YOLO26-N-320 → YOLO26-N-384 → MiDaS-256 → DA-V2-196 → DA-V2-252.
Measured on-device update rate per model (still being investigated for further performance improvements):
| Model | avg inference | update rate |
|---|---|---|
| MiDaS-192 | 83.6ms | ~12.0 Hz |
| YOLO26-N-256 | 85.0ms | ~11.8 Hz |
| YOLO26-N-320 | 132.2ms | ~7.6 Hz |
| MiDaS-256 | 152.3ms | ~6.6 Hz |
| YOLO26-N-384 | 203.5ms | ~4.9 Hz |
| DA-V2-196 | 278.8ms | ~3.6 Hz |
| DA-V2-252 | 488.0ms | ~2.1 Hz |
Depth Anything V2 is new here - it never actually worked before (the previously-bundled asset silently failed to load on-device), so this is a genuinely new option, not a fix.
Widescreen (21:9/4:3) fixes
- Fixed 21:9 and 4:3 sources being squeezed into a 16:9-shaped screen instead of displaying at their real aspect ratio (thanks to @smturtle2 for the detailed bug report)
- 21:9 is now two resolution options - 21:9 HD (2560x1080) and 21:9 2K (3440x1440) - matching the HD/2K split every other aspect already has
- Fixed a garbled horizontal band that could appear with AI 3D enabled on very wide sources
Also
- Fixed YOLO26 depth models producing visibly blockier output than they should (re-exported with a different quantization approach)
Contributors
smturtle2