search

TheLastBen / fast-stable-diffusion

fast-stable-diffusion + DreamBooth
MIT License
7.51k stars 1.31k forks source link

Is there a way to locally install fast dreambooth? #1780

Open Princess-Joanna opened 1 year ago

Princess-Joanna commented 1 year ago

Apologies if this has already been answered, but every link I could find on it is no longer up. I was wondering if there's a way to use fast dreambooth locally? I have 24GB vram, so I prefer to use dreambooth locally rather than worry about Google Colab limitations and having to pay to avoid interruptions.

I've used other dreambooth repos, but this one by TheLastBen is the one that gave me the best results closest to the look and style I was going for, so if possible I'd like to keep using it but offline.

Thanks!

HorrySheet commented 1 year ago

I could be wrong but isn't it an extension downloadable from within the Automatic 1111 GUI that you should be running locally to begin with? I'm sure others can answer you better but I think you'll find what you're looking for in a youtube tutorial video search on "How to install Dreambooth Locally using Automatic 1111". Good luck :)

Princess-Joanna commented 1 year ago

I could be wrong but isn't it an extension downloadable from within the Automatic 1111 GUI that you should be running locally to begin with? I'm sure others can answer you better but I think you'll find what you're looking for in a youtube tutorial video search on "How to install Dreambooth Locally using Automatic 1111". Good luck :)

I'm still learning myself, though from what I gather there are different dreambooth repos and the one used in the fast-dreambooth colab is different to the Automatic 1111 dreambooth extension. I could be wrong there, and they could be exactly the same, but just from using the two with similar settings TheLastBen's one always looks better imo.

From the posts I've seen around fast-dreambooth I assumed this was a modification of the existing dreambooth, which is why it always looks cleaner with better results when using this version, but if it's not then maybe it's just a settings thing. I would need someone who knows more about it to clarify, as I'm not 100% sure myself.

HorrySheet commented 1 year ago

This is what I used to do my face in a model. I think I watched a youtube video on how to do it or just followed the prompts , it only took like 40 minutes with 22 or so pics.

https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb?

image

Princess-Joanna commented 1 year ago

This is what I used to do my face in a model. I think I watched a youtube video on how to do it or just followed the prompts, it only took like 40 minutes with 22 or so pics.

https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb?

image

I've made several models using the collab, and 2 using automatic, I haven't used kohya_ss for dreambooth specifically, but I have for LoRa. I also haven't tried the older way via the Joe Penna repo. Of the ones I have tried, TheLastBen's colab always looks the best in contrast. I thought it was based on things specific to his version of dreambooth beyond just user face settings (which I try to keep as close as possible, same with the test datasets I've been using). As I said, it could be that his repo isn't actually different and there's something else I've not taken into consideration. I know dreambooth repos differ in quality, even if they do the same thing based on how they're optimized. So I was wondering if they are 1:1 the same or not. Based on his Reddit posts, I thought they were not, but I might have misunderstood what he meant.

TheLastBen commented 1 year ago

my colab separates the text encoder training and the unet, it trains the text encoder first with specific steps and lr, then the unet with its own settings too

Princess-Joanna commented 1 year ago

my colab separates the text encoder training and the unet, it trains the text encoder first with specific steps and lr, then the unet with its own settings too

Ah, thanks for the info! Is there a way to replicate the exact same results that I get when using your colab, but locally? Or does it not work like that (I'm still new to dreambooth, sorry if this is a dumb question)

Ahmet-Dedeler commented 11 months ago

@Princess-Joanna were you able to figure out how to run it locally? the colab interruptions make it very hard to work with this 😭

Princess-Joanna commented 11 months ago

@Princess-Joanna were you able to figure out how to run it locally? the colab interruptions make it very hard to work with this 😭

I ended up following Aitrepreneur's YouTube tutorials for training SDXL models on Kohya SS, once you have your parameter's set up correctly you can save it as a config and quickly load it every time you want to do a training test. used one of the many photoreal model's as a base and have been testing different datasets. Been getting some of the best results I've been able to get so far with this approach though it can still be hit and miss depending on how I've compiled the data set (still learning!)