Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

What's the best way to use SD (3 or 2) online? I can't run it on my PC and I want to do some experiments to generate assets for a POC videogame I'm working on. I pay MidJOurney and I woulnd't mind pay something like 5 or 10 dollars per month to experiment with SD, but I can't find anything.


I used Rundiffusion for a while before I bought a 4090, and I thought their service was pretty nice. You pay for time on a system of whatever size you choose, with whatever tool/interface you select. I think it's worth tossing a few bucks into it to try it out.


Eh, you can get the same software up and running in less than 15-20 minutes on an EC2 GPU instance for about half the hourly-rated pricing of rundiffusion. And you'll also pay less than their 'premium' monthly fee for storage of keeping an instance in the Stopped state the entire month.

I used rundiffusion to play around with a bunch of different open source software quickly and easily with pre-downloaded models after getting annoyed at my laptop GPU. But once I settled on one particular implementation and started spending a lot of time in it, it no longer made sense to repeatedly pay every hour for an initial ease-of-setup.

The only real ongoing benefit was rundiffusion came with a bunch of models pre-downloaded so swapping between them was quick. But you can use UI addons like the CivitAI browser to download models automatically through automatic1111, and you'll likely want to go beyond what they predownload to the instance for you anyway.

The downside to running on the cloud directly is having to manage the running/stopped state of the instance yourself. I haven't ever left it running when I was done with an instance, but I could see that as a risk. CLI commands and scripting can make that faster than logging into a website which does it for you automatically, but it's extra effort.

I thought about building an AMI and putting it up on AWS marketplace, but it looks like there are a few options for that already. I don't know how good they are out of the box, as I haven't used them. But if spending 20 minutes once to get software running on a Linux instance is truly the only barrier to reducing cost, those prebuilt AMIs are a decent intermediary step. They're about $0.10/hour on top of server costs. I skipped straight to installing the software myself, but even an extra $0.10/hour overhead would be better than paying double..


Would you recommend that to someone who has never used AWS before? Is it possible to screw up and rack up a huge bill? I might consider using that for big tasks that I can't do with my local setup.


It's a _little_ possible to generate a huge bill, but the biggest risks here are:

1. Leaving instances running when they're not being used

and

2. Deviation from default behavior that results in accumulation of storage volumes you don't want or need (low likelihood but something to watch for initially).

For 1:

If you leave the instance running you'll keep getting charged the hourly rate. Not really unexpected, but you have to notice it yourself or set an alarm.

There are a few tricks to reduce likelihood of this happening and to limit charges if it does happen anyway:

a. Prevention: Make your own little auto-stop script for the instance like rundiffusion has. Maybe make it into the launch sequence too, so you run a script, it launches the instance, then starts a timer. If the timer counts down all the way with you jiggling it, it stops the instance.

b. Mitigation: Create an alarm on the instance with the action to 'Stop' the instance when the alarm is triggered. Set the trigger for the alarm to be something like 'Max CPU usage has been less than 4% for a consecutive hour'.

c. Mitigation: Use AWS' Instance scheduler to automatically stop the instance

d. Mitigation: Billing budgets with associated action to stop instances -- kind of like the alarms but triggered based on costs

For 2:

It's probably a non-issue. You'll likely not have a problem because you'll start and stop the same instance most of the time instead of creating and deleting new instances. In which case, gp3 SSD storage is $0.08/gb over a month, charges on a 200gb storage volume you keep around all the time and use is only like $16 for the month. There are benefits so it's likely worthwhile.

BUT, be careful if you create and terminate lots of instances instead of stopping and starting the same instance. There's a small possibility of accumulating extra storage volumes you don't need, without realizing it.

By DEFAULT that's not a problem. AWS will delete the attached storage volumes for an instance after you Terminate (not stop) the instance. The problem comes from changing the default behavior.

That change can be configured in the AMI you use to launch an instance (by whoever created the AMI), or by you when launching an instance. The storage volumes are not deleted automatically then you have to do it manually to stop them from continuing to generate charges. Which you might not notice right away..

Keep that in mind, but that scenario requires a fairly unlikely chain of requirements to get to the point of bill bloat:

- If you use a pre-built AMI from marketplace (such as one with stablediffusion preinstalled) which is configured to KEEP storage volumes upon termination instead of using AWS default of deleting them,

- and if you do a 'Create' and 'Terminate' instead of 'Start/Stop' so you're using a LOT of instances instead of a few,

- and if you don't notice the setting during launch,

- and if you don't see all the extra volumes sitting around...


poke around stablediffusion.fr and trending public huggingface spaces




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: