System:
Windows 10 19045.5737
Intel i9 14900K CPU
RTX 5090
64 GB RAM
Nvidia drivers 576.02
Audio2Face 2023.2.0
I never had a problem running Audio2Face, until I replaced my RTX 4090 with an RTX 5090. I now get the error “Failed to build TensorRT engine” in the interface.
I’ve uninstalled and reinstalled Audio2Face, to no avail. I have followed some suggestions about creating a new scene and adding Head Template, A2F Core etc. This does not solve the issue either.
I have Audio2Face running well with an RTX 3080 Ti on another PC. Is the RTX 5090 a compatible device?
The issue here is that audio 2 face is a very old program which has not been updated in a long time, and you are trying to pair that with a brand new video card with brand new drivers. The GPU certainly has the hardware available to do the job, but as with any video card, it’s only as good as the drivers. Nobody here has tested at 5090 with audio to face because it’s discontinued.
My advice is you can try some old drivers. You could go back to the recommended drivers a 573.70. Or you can use your 3090/4090 in a separate machine just for A2F.
We no longer support the program here on the forums. If you need federal assistance, you should go our Nvidia developer discord channel.
Thank you. Your reply was very helpful. The 4090 moved to a different machine, so I expect we’ll use whatever machine its in while I look for solutions. There appears to be a workaround on the Discord channel you pointed me to for the 5090.
Hi there. My ‘solution’ was to abandon the Omniverse Audio2Face application and migrate to their Audio2Face-3D NIM microservice, which can be deployed via Docker on your local PC. The communications protocol is gRPC instead of REST. I just modified the example application with minimal effort. The blendshapes file created is an CSV, comma-delimited, type instead of JSON. The performance may be better than the Ominverse application, too.
okay thank you for the answer! One important thing for me is that i want to export as USD file to use in 3D software. Do you know if that is also possible in the NIM version? I have read that it might not be.