SDXL Local LORA Training Guide: Unlimited AI Images of Yourself

This guide contains everything you need to train your own LORA or Low Rank Adaptation model for Stable Diffusion XL (SDXL) using your home PC. This guide will allow you to train SDXL to generate images of yourself, or anyone else for that matter.

Parameters: scale_parameter=False relative_step=False warmup_init=False

--- If you get an error : "No data found. Please verify arguments"

--- Click the "Prepare training data" button on the Dataset Preparation tab.
---------------------------------------------------------------
--- if you get an error: Creating venv...
Python was not found; run without arguments to install from the Microsoft Store, or disable this shortcut from Settings Manage App Execution Aliases.

--- it is because you need to check the box "add python to system path" when you install python
---------------------------------------------------------------------------
--- The following folders do not match the required pattern number_text:

--- If you click on the TOOLS tabs and fill in the info there (location dataset images, reg images if you have any and then enter a folder in the DESTINATION TRAINING DIRECTORY box and click PREPARE TRAINING DATA it'll create the right structure and move all the files into there for you. Then click COPY INFO TO FOLDERS TAB to make sure you have the right info in the right place.

[frankschannel]

One of the better YT videos on the topic. Liked and subbed.

[liorash]

For some reason, when choosing the sd xl 1.0 base checkpoint, I can't read from the UI the Lora files (which they're .safetensors file). With the other default checkpoint of webui, I could load them. Is this something regarding incompatibility with sd xl 1.0 base and Lora's files versions or something?

[adriandixon]

Definitely one of the better tutorials than most out there for sure, thank you! Would the process be the same for training a SD 1.5 LORA? If not then do you have another video/guide I can follow? Thank you again in advance!

[Villa-Games]

It's hard to believe that I used to delete other Loras without thoroughly checking them. And bitch around about having a poor outcome!
Thanks buddy, quite helpful

[freneticfilms]

You are saying once you use Buckets, the resolutions dont matter and also not if they are not cropped perfectly. Fine, but doesnt it say something about "max bucket resolution" in Kohya? And will i benefit in any way from a highe Res image, that i leave on a higher resolution?

[vatoko]

Thank you very much! Very useful.
Nvidia 3060 does 10 epochs in 10 hours with Network Rank 32 and Network Alpha 16 and 25 - 42 1024x1024 images in the dataset. It's a long time, but it manages.

[RSV]

Thank you very much, a great video. Finally a more detailed description. I'm going to try it even though I only have a 3050ti/4GB vram, but it seems that the computer also uses normal memory to complement the lack of vram, although it takes longer that way.

[ICHRISTER]

Thanks for the video! I followed up with all the steps, now waiting to see the results!
Is there anything that has changed since the video was released? should I enable more options on a 3090? I currently at 19Gb Vram, some are spare.

[felipeprato]

Thank you for your proactive response to subscribers... I'll try your method soon... but I'd like to know if there's a way to make a personalized "checkponit" with personal photos for Comfy to use... I have several photos of places in mine city ​​you would like to use as a base; Can you point me to a video? How are "checkpoints" created?

[sergetheijspartner]

Also I want to make a lora for a symbol which could be used in multiple jewelry, emblems or logo images, problem is that this symbol does not have many pics, but I do have a 2D black and white picture and I have fusion 360, could I make images of a 3D render in every angle and use that as training data? I think I just answered my own question, but hey you can use the idea for a video of you like

[dwainmorris]

Yes, there are a lot of programs out there.That can do LORAS when it comes.To faced likenesses and body types.But when it comes to detailed superhero costumes.LORAS fall short, what if I want a particular costume on a likeness?Then what do you do

[freneticfilms]

If the X/Y/Z sheet resemble the epochs you used and you always go for the once more to left side (so epoch 3 to 5), why do you even set the epochs that high in the first place? Why didnt you go for 6 epochs for example? Is there any quality loss if you go for lower epochs for the results on the left side? (not sure if i made myself clear)

[ElGalloUltimo]

Great tutorial. I will note that you seemed to have skipped a very important step: choosing the Source model under LoRA > Training

[jack_lion]

Thanks for the great video! A lot of the results that i'm getting by following your method resemble the celebrity look alike or resemble the civit prompt more-so than the images of myself. I will continue experimenting of course. But I was wondering if you had any tips to make the images resemble the subject more? Additionally, is there a way to generate multiple images with a different variation or seed for each Lora model? Also I noticed that the "No-Half VAE" option was not present in my UI.

[LaguGabut]

I Use phyton 3.11
Kohya SS not compatible with it...

Is there any chance that Kohya can run on phyton 3.11?

[zhenyiyang]

Hi everyone, there is an error when I run the setup.bat file in the kohya_ss directory ( ModuleNotFoundError: No module named 'pkg_resources' ). I have already installed Python 3.10 and visual studio 2015. Any good solutions for it? thank you very much!

[itscout]

Bought your patreon, where do I find the folder of women photos as advertised in the video? I can't seem to locate it.

[andrewarc]

i did everything accordingly and whenever i get to BLIP captioning, select directory and name prefix, after pressing captions it doesn't do anything. can yall help me with this?

[Asaghon]

Nice vid, very similar to the way I train in 1.5. (except wit prodigy). But no min SNR gamma and Noise offset?

[iiJDSii]

Amazing guide, subscribed! Thank you. I was wondering if you have a related tutorial about inserting a person's LoRA into photorealistic scenes, like on a boat, mountain climbing, at a party, doing other cool stuff, etc. That would be really neat, could be an add-on to this! Cheers