search

sstzal / DFRF

[ECCV2022] The implementation for "Learning Dynamic Facial Radiance Fields for Few-Shot Talking Head Synthesis".
MIT License
335 stars 40 forks source link

Questions about run.sh being very slow #23

Open hylq66 opened 1 year ago

hylq66 commented 1 year ago

It is an excellent work, there are some questions. This is the log of run.sh training, is it normal? Is it too slow?I used your default values for all my parameters. I'm using the 0 video. image

cucdengjunli commented 1 year ago

same

kjhgfdsaas commented 1 year ago

Same. I'm using 3090. I have double checked the usage of CPU, RAM, GPU resources, RAM is low, CPU and GPU are almost full! I feel something wrong but I can't read this with my patience. Incredible bloody code!

sstzal commented 1 year ago

It is an excellent work, there are some questions. This is the log of run.sh training, is it normal? Is it too slow?I used your default values for all my parameters. I'm using the 0 video. image

The training time for each iteration is slow, which is indeed a weakness of this paper. You can reduce the number of the sample nodes in each ray, or reduce the number of the reference images to acceleration training. Moreover, you just need to train for short 1w-4w iterations to get satisfactory results.