Software Engineering Courses (SECourses)•3y ago

Drop your model there and let it upload.. might take a while since Jupyter is slow

Iidmon Drop your model there and let it upload.. might take a while since Jupyter is sl...

Longshao•3/25/23, 7:45 PM

Thanks, I'm new to this, but I'll follow your instructions, do you know of any video tutorials etc?

idmonOP•3/25/23, 7:46 PM

I think Furkan made some excellent tutorials explaining Runpod and Jupyter

idmonOP•3/25/23, 7:46 PM

problem is this space moves so fast. A tutorial from last month might be outdated today

Longshao•3/25/23, 7:46 PM

Facts, lol

Rewire•3/25/23, 7:48 PM

I cant get a good model out of dreambooth, and training takes hours on my 3090. Also, it keeps trying to make thousands of class samples.

What gives?

idmonOP•3/25/23, 7:49 PM

I'm training dreambooth as well on my 3090... it shouldn't take hours

idmonOP•3/25/23, 7:51 PM

Did you provide a folder with pre-generated class images?

idmonOP•3/25/23, 7:51 PM

Because if you don't provide it... it will generate it based on this value

brenkhof•3/25/23, 7:59 PM

Does anybody have a good fresh tutorial on how to install dreambooth on runpod? I used Furkans last one, but that didn't work

TMorgan•3/25/23, 8:51 PM

@Furkan Gözükara SECourses Hello! I finished downloading all the ControlNet files and am continuing to get the FileNotFound message. I could not find the cldm_v15.yaml file, either.

Iidmon Did you provide a folder with pre-generated class images?

Rewire•3/25/23, 9:19 PM

Every tutorial I've watched says nothing about providing class images. They always just seem to generate them?

Rewire•3/25/23, 9:20 PM

Picture of a man

Rewire•3/25/23, 9:20 PM

For example

idmonOP•3/25/23, 9:20 PM

You can generate them yourself and save them for future trainings

idmonOP•3/25/23, 9:21 PM

Store them in a folder "man", "woman", "..." all the classes you frequently train

idmonOP•3/25/23, 9:21 PM

and link back to that

idmonOP•3/25/23, 9:21 PM

saves time

Rewire•3/25/23, 9:21 PM

Interesting

Rewire•3/25/23, 9:21 PM

I mean after the training I can generate pictures of my face just fine but attaching them to something else like a superhero costume doesn't do anything anytime I insert my instance prompt it just generates a picture of my face it won't ever do anything extra beyond that.

Rewire•3/25/23, 9:23 PM

Can I use my own preprocessed images with clip descriptions to act as a class?

Rewire•3/25/23, 9:23 PM

Or will that skew the training data.

idmonOP•3/25/23, 9:23 PM

Yes I do that myself actually

idmonOP•3/25/23, 9:24 PM

I generate them on the same model I train on (important detail) on sampler: DDIM (default diffusion method for training) on 30 - 50 steps

Rewire•3/25/23, 9:24 PM

Interesting. Any luck with adding additional prompts though like making yourself into a superhero or making yourself into like iron Man or something.

Rewire•3/25/23, 9:25 PM

I just can't seem to get past the step of generating my own face it does that just fine but I can't attach it to anything.

Rewire•3/25/23, 9:25 PM

Wondering if there's an issue with my workflow.

idmonOP•3/25/23, 9:25 PM

I usually try to generate class images that look very similair to my subject. I try to write detailed prompts and do a little bit of "class bleeding"

idmonOP•3/25/23, 9:26 PM

It can help with converging your model faster using class images

Rewire•3/25/23, 9:26 PM

Hmm

idmonOP•3/25/23, 9:26 PM

Rewire•3/25/23, 9:26 PM

I'll try that when I get home.

idmonOP•3/25/23, 9:26 PM

instead of doing just "man"

idmonOP•3/25/23, 9:27 PM

I went for specific key features

idmonOP•3/25/23, 9:27 PM

still experimenting, but so far helped with this specific dataset

Rewire•3/25/23, 9:27 PM

Yes I see now. I guess I'll have to fiddle with it more when I get back home.

Rewire•3/25/23, 9:28 PM

I'm just trying to get myself looking like I'm sitting at a high-end bar in a business suit drinking a gin and tonic or something. But it just generates pictures of my face after I do the dream booth training which is effectively useless.

Rewire•3/25/23, 9:28 PM

I feel like my prompts are pretty detailed but whenever I add my personal instance prompt tag for my training it just generates pictures of my face and doesn't do any outpainting or additional content.

idmonOP•3/25/23, 9:29 PM

Then it's overtrained, the UNet is overcooked

idmonOP•3/25/23, 9:29 PM

How many steps did you train?

idmonOP•3/25/23, 9:29 PM

How many images did you had?

Rewire•3/25/23, 9:31 PM

Oh well then that makes sense yeah I mean whenever I follow the tutorials online for training or use the built in "wizard" I end up getting to like 12,000 training steps