mpaepper

I do both, I generate one image on demand and query others from lexica :)

Author here:

I wanted to print a stable diffusion design on a t-shirt as a birthday gift, but needed to solve technical challenges to do that (quality, upscaling etc). Then I thought it'd be fun to provide this as a service for others.

You simply enter your design wish as a text and via stable diffusion a design is generated for you and you can see how it looks on a t-shirt.

Ask me any questions!

Here a bit more in depth how it works: https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

Congrats on the launch, it looks very nice and works well. And kind of funny as it was just the entry above my own Show HN where I featured creating and printing t-shirts using stable diffusion, so quite related :D

Do you use ESRGAN for upscaling the images for print as well?

I am currently mostly using ESRGAN for upscaling

What AI model do you use to scale the images up for printing?

Author here:

I built this as a side project when I wanted to print a stable diffusion design on a t-shirt as a birthday gift. The challenge is that the stable diffusion images are only 512x512px, so I use other AI models to upscale them to print resolution.

On the site you can directly experiment, i.e. enter a prompt -> get a design illustrated on a t-shirt.

Alternatively, you can upload existing designs, for example from https://lexica.art

Have fun and let me know what you think and how I can improve it :)

Here is an animation of how it works: https://www.thisshirtexists.com/wp-content/uploads/2022/09/t...

I wrote some of that as a blog entry: https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

How do you scale the models on the GPUs? Do you rent cloud GPUs?

Here is a guide for AMD, but I don't have such a card, so haven't tried.

https://rentry.org/sdamd

Looks crazy and comfy at the same time. Do you happen to know the price range?

Most GPUs are non-deterministic - learned this the hard way in deep learning on pathology data. This is for optimization purposes. In fact, you can set a flag in Pytorch / Cuda to disable this which comes at the cost of performance.

Check out this demo here where you can draw and then use your drawing and a prompt to generate new images:

https://huggingface.co/spaces/huggingface/diffuse-the-rest

To understand how stable diffusion works, see:

https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

Shameless plug if you want to understand how stable diffusion actually works:

https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

Check here for details about stable diffusion: https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

This also very much depends on the inference use case / context. For example, I work in deep learning on digital pathology where images can be up to 100000x100000pixels in size and inference needs GPUs as it's just way too slow otherwise.

If you want to read an enthusiastic, but less exaggerating piece which explains how the tech works: https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

If you want to learn how and why stable diffusion works: https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

To understand the more technical aspects of how it works: https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

For more details on how it works: https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

Alternatively: "Mosquitos hate this trick"

Nice, looks like they pretty much automated what I described in my article as a service.

Great question, I haven't tested this, yet.

One thing I'm sure would work if the user has generated two separate public-private key pairs and you sign two different certs.

Not so sure about having several certs for the same key.

If I was in that situation, I would probably generate a new cert which contains the concurrent claims and is short-lived, but we also don't have extremely many different roles.

Of course, if your CA is taken over, you are in big trouble.

However, the attack vector here is much lower given you can have this on a local machine which is not even connected to a network if you like.

The single point of failure for central management solutions is in having them as a running service - if that goes down, you are in trouble.

And that is far more likely than your CA becoming compromised somehow.

Also in practice, you can also run two or more CAs (we do that), so you can deprecate a full CA if the need arises.

Thank you, I agree that this should be avoided as much as possible in general.

Actually, in my company, we are pushing the logs to an external service, so I guess that line I wrote there was not totally thought out. ;)

However, it still might be useful for certain people to check servers for health or other metrics and you might not always have everything as perfect as you wish. I guess you already stated that yourself.

So in general I'd agree. Good comment!

What about having different access roles for certain people to certain servers?

Is that covered in the NixOS approach as well?

Sounds interesting!

The certificates have an expiration date, so if something goes wrong on that end at least it will expire after some time. Also, this takes care of role-based access - this you might not have with an authorized_keys file solution?

It's the one I submitted: https://www.paepper.com/blog/posts/how-to-properly-manage-ss...

A while ago, I asked here: "Ask HN: What do you use for SSH key management of teams?" (https://news.ycombinator.com/item?id=24157180)

And in this blog entry I am summarizing what I learned and what I think is a very good approach now.

I am, but my question is how to control for people still adding keys in addition to having logged in with a certificate.