OPTIMIZE FACESWAP REPO ON AWS (ROOP-UNLEASHED)

I am adding video faceswapping to my Electron app software.
I am using (link removed) for the faceswapping.
It has great docker support making it easy to implement.
On my AWS, I already have the service setup on EC2 with auto scaling groups with a [login to view URL] instance.
The roop-unleashed is a public repo that already has a built in API system with Gradio and the EC2 instance is connected with an API Gateway & AWS Lambda for making API calls.
In my Electron app, we provide an input face and a target file and this service successfuly faceswaps the video through the API and returns the generated file in an output folder.
IMPORTANT.
Although I already have the service setup, I am considering completely changing everything because it isn't cost effective.
My main issue is that AWS charges by the hour.
So if one person decides to run only 1 single faceswap, first they have an annoying 2-3 minute cold start for the instance, and then they faceswap the file.
And if they only generate 1 file, we got billed a full hour which is $0.
for g6.
large.
Understanding my customer base is important for this project.
I have around 20 people using my software.
It's mostly marketing agencies that will need to faceswap short form content for Instagram/TikTok.
So the max video length is gonna be 15 seconds and just a few MB max.
I also don't expect more than videos to be generated per month, and if possible, I would like to keep a budget of $/mo.
And with the current system, I could easily exceed $200 as it isn't the most effective.
So I am mostly trying to explore ways to save costs, the main thing I am trying to explore is only paying for the COMPUTING THAT WE ACTUALLY USE, and not being billed by the hour.
I am not deeply educated in AWS, but some things that I considered are AWS Sagemaker, AWS ECS with Fargate.
I would like to know if this would be the right fit for my needs.
Anything that's gonna allow us to quickly generate videos, preferably avoid 2-3 minute cold starts and only pay for the computing that we use.
And since my software only has around 20 users, the faceswap won't be spammed or expect crazy amounts of usage as mentioned earlier.
I don't know too much about the 2 methods I mentioned, so I hope to get more insights & professional opinions from someone more educated than me.
And at the end of the day, the system I currently have might already be good but just needs better optimization.
Also keep in mind since we are faceswapping, GPU matters, CPU only instances can't perform well.
So I am looking for someone to provide me with an effective solution for my problem & estimate the costs for me with the provided solution.
Previous experience with faceswapping is preferred, in your application, let me know if you have it and how you would approach my issue.
Check out the github repo, it's ready to be dockerized and doesn't require any model training.
IMPORTANT.
I am looking for the best way to perform faceswaps at any time of day and only pay for the computing that we use and avoid cold starts.
Is it possible to achieve that, if so, what is the best way to achieve it? Thanks! Amazon Web Services Python Docker Git Machine Learning (ML) ID do Projeto.
# Sobre o projeto 19 propostas Aberto para ofertas Projeto remoto Ativo em 11 minutos atrás