Toggle each effect independently to see exactly what it contributes. Everything runs in your browser; the AI models run once when the photo loads, then the 30fps loop just composites layers.
Effects 1, 2, 4 need only the segmentation model (~250 KB). Effect 3 additionally loads the pose model (~5 MB) — it is fetched lazily the first time you tick box 3.
Parallax reveals whatever was behind her, which no photo contains — so the gap is patched with a blurred stretch of the background. Keep strength modest on busy backgrounds or the patch becomes visible.