Compute details

Thanks for sharing your idea! From the paper, I understand that your model can run in real time on a CPU. However, when I looked into common lip-embedding model architectures (as mentioned in the paper), I noticed they typically use a combination of 3D convolutions and ResNet-18. Estimating the computational cost, the MACs seem to exceed 12G at 25 FPS (only for lip embedding generation). Could you please confirm if this is accurate?

If so, would that imply the model can achieve low-latency performance only on high-end CPUs? My understanding is that standard CPUs, with clock speeds around ~2 GHz, might lead to higher latency for such workloads.

Start by reviewing the paper's real-time CPU claim and the referenced lip-embedding architectures, especially 3D convolutions and ResNet-18. Verify whether the estimated 12G MACs at 25 FPS is accurate and document the resulting CPU and latency assumptions; done means the compute and low-latency claim are clearly confirmed or qualified.

Written by the indexing model from the issue text.