1000's SDXL REGULARIZATION IMAGES
🕑 Added 2023-08-11 17:24:31 +0000 UTCI've meticulously curated and resized thousands of high-quality regularization images for various class prompts for SDXL training, and I'm offering them exclusively to my supporters!
What's included?
- 1000s of images of 'Man'
- 1000s of images of 'Woman'
- 2000 images of 'Person' (a blend of 'Man' and 'Woman')
- 1000s of images of 'Artstyle'
All of these are available in two resolutions: 1024x1024 and 768x768.
🤔 But first, what are Regularization Images?
Regularization images are more than just good looking images. They play a critical role in image processing and inversion.
1. Preventing Overfitting & Preserving Classes:
When we're training models, there's always a risk of them becoming too familiar with our training set, causing them to lose their versatility. This phenomenon is known as overfitting. Regularization images help prevent that.
Moreover, let's say you're trying to train a model to recognize an airplane. Without proper guidance, it might mistake it for a car, bike, or even a toy plane. By using airplane regularization images, we're essentially giving our model a "classroom lesson" on what airplanes truly look like.
2. Enhancing Editability:
One of the challenges with training is retaining the editability of images, especially with prolonged training. My curated regularization images ensure that the images retain their uniqueness and don't become carbon copies of the training set.
These regularization images are a result of countless hours of curation and refinement. They're designed to give your training a significant edge, ensuring that your models are trained effectively while retaining their uniqueness!
LINK TO DOWNLOAD (PATREON doesn't support files over 200mb....) :
https://bit.ly/reg-images
or mega link (copy paste that full URL, patreon has trouble displaying the link)
https://mega.nz/file/SAUyxT5J#kTfnFslj-cwSfAAtcERSY2OTVfQiHgFql6bc-H3zNR4
Use Winrar to extract the file using the password: aitrepreneur
Let me know if you have a problem with the download!
As always, supporting me on Patreon allows me to keep creating helpful resources like this for the AI art community. Thank you for your support - now go train some awesome LORAs!
Comments
Aitrepreneur
for the 1.5? yup, that's pretty much it
koumac kouini
hello, to do SD training, do you have to resize the Regularization images to 512? and in the settings, uncheck SDXL and use a basic SD model? that's all ? many thanks
Aitrepreneur
it works, the link is just cutoff on mobile
Some Guy
other link works fine
Some Guy
The password for the megalink no longer works
Aitrepreneur
I updated the description
Prairie Fawkes
I think the issue that people must be having with the Extract not working and the files being corrupted is that they're not giving the extractor the right password. I was having the same issue with 7zip because when it pops up "PASSWORD" I just assumed it meant MY password. So I gave it my PC password and it then just decrypts corrupted files. It might save you time having to answer people's comments if you really make clear in this post that when the RAR Extractor asks for password you need to enter "aitrepreneur". The way the posts makes it sound it seems like that password is just for Mega. I didn't even know RAR files could be encrypted! Thanks for the resources!
Pauline Fannes
it's good, no need to reply to my message, as I really couldn't find it, I restored my PC to a lower date (I went back 2 days) and hallelujah, it works again but only refiner, base s It's decided to strike... But that's it, I'm happy, I'll finally be able to train my lora!
Pauline Fannes
hello, it took me a while to understand how to open the training files, I finally understood by reading the comments here. But I searched a lot on my own, I downloaded a lot of software to try to restore the images because I thought they were damaged. At the beginning I could use SDXL in 512X512 with my 8GB Ram GPU but after my attempts, it's impossible to use SDXL even when I reduce the size of the images very much, what I get makes no sense, I would really like go back ! Only 2 days ago everything was working fine, do you have any idea how I could do it?
Aitrepreneur
there is not issue, use winrar to extract the archive and the aitrepreneur password
Roy
Is the WW password issue fixed? Came here mainly to save myself that work :)
Aitrepreneur
Yes I've known about it just after I did my first Lora training video :)
Justus Hulmu
For your next runbod training tutorial you can put: "tail -f /workspace/logs/kohya_ss.log" in terminal and you get logging in terminal (:
Aitrepreneur
it is valid, you need to use winrar for the extraction
Zahir Kahn
The password is no longer valid
djpashka
Aitrepreneur! thank for the images. Worked for me without any issues. I'm on PC, used winrar for extraction, just need to put in the password.
Aitrepreneur
you can send me more info in dm, screenshots also if possible
robbiek
Kohya states that it finds a valid directory structure but finds no suitable images inside it. Anyone else had this issue?
Uru
It was rude of me to make my first post a complaint about something. I apologize for that. I think your videos a/should be elevated to global treasure :)
Uru
Im having the same issue with the first site... it says:An unexpected error is preventing the archive from being extracted ...Error 0x8096002A:No error descrption available. I should note that I am using winrar.
Aitrepreneur
I can assure you it works, the bit ly link sould work better than mega, use winrar to open the archive with the password
Allucinari
EDIT: I was able to open this on a mac, then transfer back to PC, but on the PC, just to point out, I could not open it: Original post: I've downloaded this twice from bit.ly and tried to open it both with system and win-rar but it seems to be corrupted as it only gives the "man" folder and no images. The mega download stops just under 5GB and requires a membership or waiting six hours.
Aitrepreneur
use winrar for extraction
Franco Antonelli
incorrect pass
shen oracle
how to use it,when I train a man,I choose which one from your pakage
jamal
i cant extract the folder. so i tried going through mega but copy and pasting the password says it is incorrect
jamal
it doesnt let me extract the files
Aitrepreneur
yes I also recommend use winrar in case it didn't work before
David Mekibel
nvm i just forgot to put the password in to the unarchiver when I was unzipping
David Mekibel
hello photos seam to be corrupted. atleast woman 1024 for sure
Aitrepreneur
you can use them for 1.5 as long as you resize them to 512x512 resolution
Robert Merrill
is the tagging structure different? Wondering if these will work on 1.5SD as well.
Xaddre
any chance you could split up the files into different MEGA files because its 5.05 GB and the MEGA free plan only supports 5 GB downloads for free plans unless I want to wait 6 hours to finish the last 50 MBs. Edit: NVM I didn't see the second download link I'm good :)
Aitrepreneur
You can try following this guide: https://www.ezyzip.com/articles/en/how-to-open-password-protected-rar-file-on-macos/
Sam Lavy
Hi, i'm on Mac, I don't have winrar. Can't seem to open it. The password is not correct. What should I do?
Aitrepreneur
maybe yes, I'll probably do another video on animals
leo kim
any plans for dog images..?
Aitrepreneur
use winrar for extracting the files
StableStab
it says the password is invalid
Aitrepreneur
You need to use them for kohya training, when it asks you for the reg folder
Edmund Tan
hi sorry- am new here- where should i place the unzipped contents?
Liam John
Can you make it so people can download each folder separately because it is too big to just be a normal folder when I download it. Thank you!
Aitrepreneur
if you are training a style I would actually recommend not using reg images, compared to training a person, their use is very optional, I will explain it in the next video
Alex Kuu
Thanks for the material. I understand that you have already discussed these topics, but I would like more precise answers from you. I have two questions. 1. I want to create LoRA for the style of grimoire pages with yellowed paper and pencil or charcoal drawings on rough paper texture. I have a question: do I need to create REGULARIZATION IMAGES of the same type as the images based on which I am training, or do I need utterly different image styles from your set? 2. Are there any rules about the number of REGULARIZATION IMAGES? Are 10 REGULARIZATION pictures enough per 100 training pictures? Or do I need 100 REGULARIZATION pictures with 10 educational ones? Or other proportions.
DTar5478
Thank you! Noob, but learning fast
Aitrepreneur
they need to be used inside kohya for lora training when it ask you for reg images
DTar5478
What folder should these images reside in? Thanks!
Vlad S
Any suggestion on how to train lora for two different people at once? So SDXL can create pictures of both together and faces will not be a blend of features that belong to both?
Mike Menders
Fantastic collection! Many thanks!
Aitrepreneur
for this you should use images of shirts, I might make a separate video on objects
Aitrepreneur
Yeah took me a loooong time too :D
Michael
which set should i use if i wanna train a object like a shirt?
Ramon van Schie
Thanks I was just looking for this! was going crazy manually scraping images.
Aitrepreneur
download winrar it works believe me
robot sound
rar file will not uncompress, gives error "unknown archive type, only plain RAR 2.0 supported(normal and solid archives), SFX and Volumes are NOT supported!"
Aitrepreneur
I made another download link no need for mega: https://bit.ly/reg-images
JoeyDotDigitize
I just saw the mega website makes you wait like 5 hours after 90% complete, but you can resume it.
JoeyDotDigitize
The mega link says you have to sign up to download the complete file. :( But I have an idea, you can just build a webserver from your home, let me see if I can build one. I had one up but had to take it down. On FIOS speeds too.
Aitrepreneur
they did have reg images during training, I didn't put them in the folder since you have them right here.
Aitrepreneur
really? It usually takes a few seconds to upload, it's like 100mb of images total
Aitrepreneur
I found them on websites like Unsplash and other copyright free images sites, compiled them, took out the ones that didn't look like the class prompt I was searching for (man-woman-style) and resized them to the right resolution. The variation in the images and poses again play an important role in making sure that the model truly understand what you are training, that the model does not drift into something that is not what you are trying to train and that the final model is not overfitting, meaning that It will start to look like the training images, so we need variation, number and quality all at once in order to get the best results, of course a few ok images are not going to influence that much the whole model, also because model reg images are not even used during the whole training, it all depends on the number of training images but the ones I provided are moooore than enough as long as you have less than 50 training images, in that case you might want more regularization images, but this is really an exceptional case, you really don't need this much images to begin with.
Boxertwin
Thanks for this great resource. Could you give a little more advice on their use? I noticed the 4 LORAs below didn't have regularization images in the dataset, so just curious when you would use these and how many in relation to the overall image set, etc....? Thanks!
Craig
Can you add these as a tar? I find runpod can be a bit hit or miss for uploading batches of files so it would be nice to use wget to download these images.
James Clow
Hi, Thank you that's brilliant. You say that "we're essentially giving our model a "classroom lesson" on what airplanes truly look like.". Where did you get the regularization images from? Also would there be any benefit to me swapping some of the lower detail images with higher detail ones, or does the pose play an equal part? I'd love to learn more about regularization images and the part they play.
Aitrepreneur
I also added another link for the download, it should work fine now ;)
Steve
Thanks worked fine! Initially I misunderstood the message mega.nz was trying to upsell I was able to click on the bottom and skip that and it downloaded in my browser wouldn't allow me to put it into a download manager...any other secrets for captioning that aren't on youtube or useful workflows to get the lora to shine on comfyui?
Bobbie
thanks for this
Aitrepreneur
you can use megabasterd for that but I'll try to find another hosting service also
Aitrepreneur
you can use megabasterd for that but I'll try to find another hosting service also
ZanderZrantZ
Mega last a limited for free accounts and your file is right above the 5 gb limit (fwerid don't seem to have an issue with the models one but we do with this one ) not sure if u wanted to change that.. (it looks like is telling me i need to wait 6 hours to be able to contiune for free)
GabeMade
Can you please get a mirror link or upload the folders individually instead of all of them in 1 zip? Mega has transfer limit of 5gb. Its not possible to download the file without a paid account as it is 5.05gb.