ULTIMATE SDXL LORA Training! Get THE BEST RESULTS!

In this video, I'll show you how to train LORA SDXL 1.0 using YOUR OWN IMAGES! I spend hundreds of hours testing, experimenting, and hundreds of dollars in cloud computing training to bring you the ultimate LORA training guide for complete beginners and experts alike. SDXL is incredibly easy to train as long as you know what you are doing with the right training parameters. And after this video, you’ll have the required knowledge to train anything you want with SDXL 1.0 LORA with the Kohya GUI tool.

What do you think of SDXL LORA training? Let me know in the comments!
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
SOCIAL MEDIA LINKS!
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Jup1t3R!
scale_parameter=False relative_step=False warmup_init=False
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
►► My PC & Favorite Gear:
Recording Gear:
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Special thanks to Royal Emperor:
- Totoro
- TNSEE

Thank you so much for your support on Patreon! You are truly a glory to behold! Your generosity is immense, and it means the world to me. Thank you for helping me keep the lights on and the content flowing. Thank you very much!

#stablediffusion #sdxl #lora #aitraining #texttoimage #imagegeneration
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
WATCH MY MOST POPULAR VIDEOS:
RECOMMENDED WATCHING - All LLM & ChatGPT Video:

RECOMMENDED WATCHING - My "Tutorial" Playlist:

Disclosure: Bear in mind that some of the links in this post are affiliate links and if you go through them to make a purchase I will earn a commission. Keep in mind that I link these companies and their products because of their quality and not because of the commission I receive from your purchases. The decision is yours, and whether or not you decide to buy something is completely up to you.
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Timestamps:
00:00:00 - Train SDXL with LORA
00:00:41 - What is LORA?
00:01:20 - Koya SS GUI Installation
00:04:36 - Koya SS GUI Launch
00:05:29 - Image Dataset preparation
00:08:20 - SDXL LORA Training
00:09:28 - LORA Folder & Training Options
00:10:22 - Character Training Tips
00:12:54 - StarByFace & Celebrities Training
00:14:01 - SDXL Character Input
00:17:15 - Training Data & Image Captioning
00:19:44 - Kohya SS & Captioning Tips
00:21:58 - The right Training Parameters
00:24:33 - High batch size Model Comparisons
00:26:01 - Learning Rate for SDXL LORA
00:27:36 - Training Tips & LORA SDXL
00:28:31 - SDXL Image Resolution & Buckets
00:30:09 - LORA Network Rank & Network alpha
00:31:47 - Image Quality & Settings
00:32:53 - Training & VRAM Tips
00:34:37 - GPU Solutions & RunPod
00:36:09 - RunPod Setup for LORA Training
00:37:30 - Training Prep & Kohya Overview
00:39:43 - LORA Config & Data Setup
00:41:27 - Model & Captioning Setup
00:43:27 - Training Log & Model Saving
00:44:45 - Model Transfers & LORA Access
00:45:46 - Image Gen Settings & Tips
00:48:01 - How to select the right Model?
00:49:04 - Image Enhancement & Output
00:50:14 - Model Comparison & Saving
00:51:17 - Conclusion on LORA Training

[Aitrepreneur]

HELLO HUMANS! Thank you for watching & do NOT forget to LIKE and SUBSCRIBE For More Ai Updates. Thx <3
"K" - Your Ai Overlord

[mactheo]

Give a man a Lora, and you feed him for a day. Teach a man to train Lora, and you feed him for a lifetime. Much appreciated K! <3

[hanmin]

An updated guide to this tutorial would be amazing as many things changed in the kohya ui and its very confusing! Thank you for all the hard work ♥

[MysteryGuitarMan]

Thank you @Aitrepreneur! I love that this tutorial dispels some myths about LoRAs. Especially the random token thing... starting all the way back from "sks", now to "omhw" – when you take Lensa and other apps like that into account, think of how many millions of GPU-hours have been wasted (they could have started from "person" or "portrait"). Only one thing to mention:

You don't need regularization images unless you plan on merging in your LoRA into your checkpoint. Or some other pretty specific use cases, like de-overfitting a specific person / character / etc.

That should speed up your training even more.

[TransformXRED]

Edit : chapters are here now
----
Don't get me wrong... I'm very grateful for your videos.

But you need to add chapters, especially for long videos like that.

People will come back more easily to it multiple time to check the tutorial... Double win.

[kallamamran]

Great video! As allways, BUT.... Focus is still on persons. This is greatly limiting since most models are allready great at creating images of persons, specific or non-specific. What I miss is a training video on how to train styles like "my own art style" or poses like "yoga/contorted/laying down poses" or maybe actions like "playing football/fishing/linedancing" and such... Just training a person (portrait/likeness) is what everyone and her mother has been doing since training has been introduced.

[joepark]

First of all, there is no sufficient explanation on configuration files. It seems the only way to get them, is to be your patron. Your customized configuration files for your patrons, well that's okay. But this whole video could have been much more useful if you mentioned where to get other configuration files, or at least how to make one. Second, the gui version you are using must be outdated, -or-, it must have some addon that you didn't mention. I've just done a fresh install of Kohya ss and THERE IS NO DEPRICATED TAB UNDER LORA - TOOLS!!! And I've just found out that in my GUI the tab is named, "Dataset Preparation."

[amin]

Hey it would be nice to see a style training guide for SDXL 1.0

[MarcSpctr]

although, SD team are right about training being faster and easier with regular tokens rather than RANDOM TOKENS, it becomes useless if you want to use the trained Loras on different base model.

So say tomorrow RealisticVision or similar base model is released for SDXL, using these Loras will result in inferior quality as compared to Loras that are trained FROM SCRATCH.

So I would suggest if you plan to use other Base Models (which ofc everyone does), use RANDOM TOKENS like ohwx, ab12 or anything random stuff.

[testales]

The deprecated section is probably labeled that way because training with regularization images is more or less obsolete or has very specific use cases only. The model already has learned millions of things and proably can take a few images more. For concepts you may even be unable to generate regularization images in first place because the concept is not yet known. By overriding training of a celebrity you are damaging the model intentionally which regularizitation is supposed to prevent. But because the Lora is applied only temporary this doesn't matter anyway.

[elias]

Wow this video did not feel like ~1 hour - thanks for making such a comprehensive guide K!

[TheKuzmann]

From my experience captions should be used in the following situations and in the following manner: use them in cases where you want to generate a specific scene, subject or concept. This of course depends on the dataset you're training on - if you're training an item, you need the dataset to consist only of that item with different backgrounds. If you are training a person's face or half body and want to generate images of that person, for example, dancing, training a model with captions that do not mention a person is dancing (or standing in a pose that implies movement, so the captions are written with the mention of it's hands in the air, describing a movement, etc.) will make it much more difficult or impossible for the model to generate a trained person dancing. On the other hand, if your dataset consists of images of a person dancing, using captions will make a desired "concept" (implying a certain person dancing, i.e. standing in a pose that implies movement etc) become a variable (I've seen that for it also a term "pruned" caption is used) which is easy to call up. On the other hand, in terms of style: training text encoder is undesirable, because you want to transfer a visual identity to the model, and most importantly, you want it to be "printed" on every possible prompt. In that case, only the class (style or aesthetics) is trained. The most common mistake is to train a style with regularization images plus text encoder (which I did for an absurdly long time training styles in dreambooth). Such a model is literally unusable and generates random images. Even training textual inversion for style using captions can make it less flexibile. I'm writng all this from my personal experience and from all the possible tutorials that exist on internet and YouTube, and I've gone through ALL of them - including yours :-) I can't even mention how many failed models I've trained, and that's necessary to learn how to train a neural network.

[juggz]

@Aitrepreneur I just wanted to point out 2 settings that it seems you may have misunderstood. At around @30:00 you mention that the "cache text encoder outputs" option is broken and suggest not to use it for now, then later at @32:50 you mention the parameter "--network_train_unet_only" and how the difference is negligible and suggest people not to use it either, BUT if you use the "--network_train_unet_only" command it fixes the "cache text encoder outputs" command. Together they use significantly less vram and makes training much faster. So the difference is negligible and the training is way faster if you use them in combination. Give it a try and you may recommend the opposite of your conclusions testing them separately.

[abdelhakkhalil]

You know, you can keep the less trained LoRa as your main LoRA and use the more trained one in the positive prompt for ADetailer. This way, you get both flexibility and details.

[OriBengal]

Glad to see you doing visual stuff again, not just LLM's. I support you pursuing all your passions, of course - but you were one of the best at creating really useful visual tutorials.

[mihoilo]

Can you please make an updated guide to creating lore for SDXL?

[marcelschuberth]

just a little hint since reopening the file to check the status is pretty annoying and inefficient, use tail -f <filename> instead, it prints the last 10 lines (by default, can be specified with -n <number>) of a file and -f sets the flag for it to update whenever new lines are added. it even handles progress bars correctly instead of printing a new line for each update

[TheRemarkableN]

You are doing the AI gods’ work 🙏. Thank you good sir. You also have excellent taste in celebrities.

[BetzVRz]

Best results I have gotten ever with this tutorial Amazing stuff

[aiviistudio]

Thank you @Aitrepreneur! I really love your contents. You have very deep knowledge about what you are doing and explain them very well. Can’t wait for your checkpoint training video ☺️