search

ashawkey / RAD-NeRF

Real-time Neural Radiance Talking Portrait Synthesis via Audio-spatial Decomposition
MIT License
878 stars 153 forks source link

Real-time execution #11

Open pegahs1993 opened 1 year ago

pegahs1993 commented 1 year ago

The process of generating the video in Google Colab takes one minute, despite the fact that it is real-time. Why is this the case?

ashawkey commented 1 year ago

@tylersky1993 Real-timeness usually refers to >= 20FPS in synthesizing each image frame. Our method can reach this FPS in V100. Colab's Tesla T4 is slower, so it's about ~10FPS.

Erickrus commented 1 year ago

I ran the latest version on colab (--torso, model engm, downscale )

NVIDIA-SMI Driver Version: 525.85.12 CUDA Version: 12.0, Tesla T4 15360MiB

The performance is:

fps size downscale
12.01 it/s 512x512
20.18 it/s 256x256 2
24.25 it/s 128x128 4
sourabharsh commented 11 months ago

@Erickrus Can you please share your google colab link? I am not able to change the size using downscale.