The Lens Studio compression algorithm is based on a lookup table generated by K-means clustering. Therefore, we do expect some accuracy loss as weights are mapped to their corresponding center weight. Weight distribution can affect clustering results but most times, the drop in accuracy is negligible. Experiments show that weight decay with 1e-4 could make models more compressible.
Lens Studio Support
- How do I get started on the new Developer Forum?
- How should I report bugs and crashes?
- Can I print my Snapcode before my Lens is complete?
- Can I keep my Lenses private?
- How can I combine multiple templates?
- Can I add a face mask to a face inset?
- How do I play multiple animations?
- Can I animate objects via code within Lens Studio?
- What should I do if I'm missing a panel or section of Lens Studio?
- Why does my PNG animation sequence immediately deteriorate in quality once I upload it as an animation texture?
- Why does my imported material show as a pink and black checkerboard on my 3D object?
- How do I use Snap Camera with Zoom?
- How do I use Snap Camera with Google Hangout?
- How do I use Snap Camera with Google Meet?
- How do I use Snap Camera with Skype?