LORA training EXPLAINED for beginners

LORA training guide/tutorial so you can understand how to use the important parameters on KohyaSS.

code: "NOT4TALENT"

------------- Links used in the VIDEO ----------

------------- Social Media ----------

Make sure to subscribe if you want to learn about AI and grow with the community as we surf the AI wave :3

#aiairt #digitalart #automatic1111#stablediffusion #ai #free #tutorial #betterart #goodimages #sd #digitalart #artificialintelligence #kohyaSS #kohya #LORA #Training #LoraTraining #outpainting #img2img #dreamlook #dreamlookAI
#consistentCharacters #characters #characterdesign #personaje

0:00 intro
0:10 What we need
0:23 Install KohyaSS
1:38 Thanks to LeFourbe
1:54 What are LORA
2:35 Best Datasets
4:26 How to get the images
5:06 Best Captioning
6:32 Captioning but AI POV
9:38 Captioning Example
11:00 Using BooruDatasetTagmanager
13:10 Training decisions
13:25 Choosing a model
14:02 Folder Structure making
14:32 What Regularization does
15:06 Steps and epochs Explained
16:50 Aprox recommendation
17:00 Ill use 14 steps and 6 epochs
17:32 Creating the folders
18:00 Training Parameters
19:25 Learning Rate Explained
20:40 LR scheduler
21:02 Use AdamW or AdamW8bit
21:30 Network Rank and Alpha
22:10 Resolution and Bucketing
23:10 Advanced Options
24:05 Train AI in minutes (sponsored)
26:10 Test Results
27:23 Thanks for watching :3

[xX_SweetAngel_Xx]

I wanna talk about some of the trait of the best loras I have found.

- They tagged everything. This makes it so you can use the description they gave their character without the lora, and as long as you have it installed, it will give more control because you can generate it without the lora then inpaint over the generation with the lora at full strength, and you will have the perfect image.

- Them tagging everything helps if you find that the character keeps refusing to wear a different outfit or their clothes tend to be the same color as their hair because the workflow I explained above corrects this issue.

- Always tag the clothes and zoom in on parts of it and describe those details so you can easily inpaint them later.

- Please edit data sets so that tiny details that fluctuate a lot in anime look consistent, and zoom in on some so people can easily inpaint and have them turn out great. Hair clips/pins and buttons can be a problem sometimes, so you may need to correct how they look. For instance, my Mamori Tokonome lora has her wearing a small white cat hair pin, and it was so inconsistent in every scene that it made the pin turn out very unfortunate, and now I need to go edit the data set to correct this.

-- Tag the backgrounds, or your characters will be stuck with the same backgrounds they were trained with. If you don't tag them, stable diffusion will assume the background is part of the character that isn't made to be changed.

- Get pictures with multiple sizes, or trying to upscale them will result in an all NANs error and/or out of memory error.

- Tag chest size on women so people can easily adjust the size.

- Describe their outfits well and their hair styles well so that if you want to keep the same face and change some things about the hair and clothes, you can do so with ease.

- Please do not intentionally create a lora that tries to keep a rigid structure. By this, I mean, do not try to limit the creative freedom of the person using it on purpose.

- When training a lora, keep to the same style all throughout. Do not use multiple different styles of the same character so that people can always know they will achieve consistent results. They can combine your lora with other loras and checkpoints to mess with the style. Please let them have that freedom.

[AerysBat]

I have trained a few LoRAs and found one principle to follow if your LoRA isn't coming out right: The problem is almost always your dataset, not your training settings. If the LoRA is not giving correct results, resist the temptation to fiddle with learning rate or anything like that. Instead, go back to your dataset and review your images and make sure you've tagged them properly. A single bad tag can really screw up a LoRA!

Once it's working well with the default settings, then you can try things like changing the scheduler, network dim to try to improve your results!

[nolan]

You explain things so well, great analogies for understanding, explained with humor and fun, and when you don't understand something you honestly admit it. Fantastic job! I'm having a lot of fun learning how to make art/comics with AI so I appreciate your videos! You've got a great teaching style brother :))

[sgl]

This is the best tutorial I found on training Loras. Finally, I trained one that works! You did a really good job on this in explaining the different parameters. Thanks!

[pokerface]

the best tutorial out there, google just recommended shit until i found this one! thanks a lot keep up the work

[pastuh]

For proper preparation of the Lora model, time is necessary.
And, if the model fails and you wish to improve it, it will take twice as much time.

In conclusion, it's a challenging task that requires dedication.
The fun may not last for an extended period of time.

[justinwhite]

Based on the provided dataset without captions, a popimpokin is a bar stool, a counter, and one or more bottles of alcohol in some combination.
(I mention this because that's all the things an AI would associate with a popimpokin without proper captions)

[mkor]

wow i searched lora video and this is the best what i ever saw keep going

[sherpya]

we definitely need a bottle of pompiko-thing

[EvilNando]

I am an absolute noob when it comes to this but I just wanted to share my experience,

so far Ive been using kohya with all default settings, just adding images croping to 512 with birme, autocaption, and hit run, 1 epoch, 100 steps, and I have been able to get the character I wanted pretty much all the time, now with this knowledge in mind i will try again properly and see how it changes

[AdrianRyoku]

Was literally looking for more lora guide videos as I've never fully understood from other guides, like hours before you uploaded. And literally after this video, I can say I am capable of actually manually making proper changes to my training parameters, whereas before, i was only using pre-made configs. Thank you again for another very very valuable educational video! I knew subscribing to you was gonna be hella worth it! Keep it up! And perhaps indulge in some popimpokin? XD

[ryry]

Been a long time since I did ai. I've only ever known hypernetworks but I don't have the patience to oversee 20, 000 steps worth of HN training.

Kinda clueless about LoRA so I hope this vid will help :D

[lefourbe]

WE MADE IT ! FINALLY !
i can take any question down bellow and i'll try to answer the best i can.
EDIT : regularization should not be used as begginer in Lora, instead use network dropout or caption dropout. this would preserve differently the model. it would speed up by two the model training.

[justinwhite]

0:04 correction on this - you can run Automatic 1111 from the CPU even if your GPU doesn't have Cuda cores nor supports Rocm.
I don't think you can train from the CPU though (haven't tried because I expect it to be unbearably slow even if it works)

[victorcorcos]

After watching this tutorial, I need a glass of Pompipokin

[piyushsonawane]

You are the G.O.A.T for making such an informative video! Really appreciate it! Thank you!

[casualgamer]

Great video! I was amazed of the sound of Poppin-pokin!

[nietzchan]

Thank you for making this, such a lifesaver~

[guilhermegamer]

OMG! Just the theme i was wodering you could choose! =D

[AnimaAwaken]

At 13:59 you have a 'deprecated' tab, I don't have that for some reason. Please help