Number Slider
Just Squares
The "newest" trend is Neural Morph Targets. Instead of an artist manually sculpting "Smile_15_degrees_left," machine learning algorithms analyze an actor's performance and generate corrective blend shapes on the fly.
| Method | VRAM | GPU time (ms) | CPU time | |--------|------|---------------|-----------| | Old (full buffers, vertex shader) | 200 MB | 2.5 ms | 0.1 ms | | Sparse deltas + compute shader | 35 MB | 0.8 ms | 0.0 ms | | PCA compressed (20 components) | 12 MB | 0.5 ms | 0.0 ms | morph target animation new
For a long piece (e.g., a cat's tail), you do not want to store a morph target for every single frame of an animation. That would be memory prohibitive. The "newest" trend is Neural Morph Targets
Instead, you use a Dimensional Reduction approach. That would be memory prohibitive
The "newest" trend is Neural Morph Targets. Instead of an artist manually sculpting "Smile_15_degrees_left," machine learning algorithms analyze an actor's performance and generate corrective blend shapes on the fly.
| Method | VRAM | GPU time (ms) | CPU time | |--------|------|---------------|-----------| | Old (full buffers, vertex shader) | 200 MB | 2.5 ms | 0.1 ms | | Sparse deltas + compute shader | 35 MB | 0.8 ms | 0.0 ms | | PCA compressed (20 components) | 12 MB | 0.5 ms | 0.0 ms |
For a long piece (e.g., a cat's tail), you do not want to store a morph target for every single frame of an animation. That would be memory prohibitive.
Instead, you use a Dimensional Reduction approach.