Fast-stable-diffusion colabs, +25% speed increase and memory efficient | Hacker News Reader

Fast-stable-diffusion colabs, +25% speed increase and memory efficient

github.com

60 pointsDLeychIC23 comments

I love StableDiffusion, and these projects are so exciting.

But. Isn't this the new equivalent of downloading and running a shell script WHICH YOU HAVE NO IDEA WHAT IT DOES? This colab notebook asks for me to give it access to my GDrive. If any of the sd code were modified from the original SD codebase (as I would expect it to be, this is a FORK afterall!) then why couldn't it pull certain anything it wants from my GDrive?

Colab does not protect you from this, right? There is no way to sandbox GDrive access, it is all or nothing, right?

Probably if you know how to run the notebook you might also know how to quickly go through the code and see if it might do a potential harm to your hard drive.

I dont see a big deal TBH.

Buuuuut.... Notice that it appears this repo contains precompiled binaries and no obvious link to the source....

This is literally running some binary blob and giving it access to a big chunk of your digital life. If it were malicious, all it would need to do is steal the token and then upload some illegal images and your whole Google account will be permabanned.

Much like you'd use a sandbox VM to test untrusted code locally, you could use a sandbox google account to work with untrusted Colabs.

Run it on a burner Gmail account

In addition to having a google account without any sensitive info, running everything in VMs locally should become a standard practice

Colab runs everything remotely unless you point it at a locally installed Jupyter runtime.

How does one make a burner Gmail account these days? Last I tried, a unique phone number was required.

Get a burner SIM card ?

if your goal is to have an empty Drive you can reuse the same number to create a new account.

Edit: i just created a new account from my smartphone without a number

There are Stable Diffusion notebooks out there that don't require GDrive access and run fine in Colab.

Everything about the SD ecosystem feels like... I'm not sure. It seems like there are tons of different small, uncoordinated efforts, all of which have either no documentation or large plaque install scripts, and then it's all python at the bottom with all of the joys of it's package ecosystem.

And then all of these versions where they're so super smart and want access to my GD and an API token to some account of mine? Christ, it would be so, so, so easy to weaponize one of these, add some screenshots and "beat" the original and harvest some Google Drive goodies.

In a, I guess I'll say, less frustrating world... there'd be a Nice or Guix expression and it would be relatively trivial to just launch it anywhere Linux and GPU exists.

Would be much better if these changes were a branch on the starting repo rather than entire refactors with many changes.

There are pros and cons to each of these Stable Diffusion hacks (usually time/memory/usability tradeoffs) so it's not correct to merge it back.

Which is why it would be a branch, no?

And in this case it seems like mostly an implementation detail - it could just be a flag to enable, or enabled whenever it detects running on colab.

did you even peek at the link ? There's a PR on diffusers, and it's mentioned on the front page https://github.com/huggingface/diffusers/pull/532#issuecomme...

These guys claim a 50% speedup with a similar approach:

https://www.reddit.com/r/StableDiffusion/comments/xmr3ic/spe...

yep, same approach but it arrived 3 days later and there's no mention of the [original PR](https://github.com/huggingface/diffusers/pull/532#issuecomme...), nice. Else the kernels used in that case -upstream flash attention- are not compatible with all nvidia GPU generations, FYI (xformers' cover a wider range and are generally faster or just pull Flash')

any differences in output or they are identical with the same prompt and seed?

identical outputs, up to float computation shenanigans (not computed in the same order, strictly speaking)

Can this be run without colab Pro? (High vram)

The un-optimized version runs in a free notebook, but you hit the vram cap if you try to increase image resolution or generate more than one image at a time.

Reply on news.ycombinator.com