Repository navigation
Has anyone actually been able to achieve real-time video generation? #396
Description
Activity
I was able to achieve realtime response on H100 GPU at 30 FPS, it used around 10-15 GB of VRAM. Your current configuration is too old and with insufficient VRAM it takes a very long time. Try upgrading your setup to 3090 / 4090 / 5090. Although you will have to play around the source code to make it real time, like increasing the batch size.
I was able to achieve realtime response on H100 GPU at 30 FPS, it used around 10-15 GB of VRAM. Your current configuration is too old and with insufficient VRAM it takes a very long time. Try upgrading your setup to 3090 / 4090 / 5090. Although you will have to play around the source code to make it real time, like increasing the batch size.
Awesome to know!
Thank you. Lately I have been using an rtx 4090 to run musetalk, and I'm getting good results after making some changes. I can get realtime inference on 15fps, 20fps struggles. a bit.
Maybe sometime I'll try an H100, although it's so expensive lol.Thats good to hear. Increase the batch size and monitor the VRAM usage. you will be able to get more than 20fps.
Reacted by AhmadAFS1Rtx 5060Ti can do about 15-20 fps or more
Reacted by AhmadAFS1
Hi all,
I'm currently testing out this musetalk library on a gtx 1060 with only 3gb VRAM.
While I am impressed with the high quality results, it takes way too long to generate a short video.
I'm waiting 1-2 hours just to generate a 20 second clip at 15 fps.
I'd like to know for those that are able to experience realtime responses, like maximum 1-2 second response time. I'd like to know what your hardware configuration is, and if you made any optimizations to the musetalk app in order to achieve faster results.
Thank you