TRAIN Perfect SDXL LoRA Clothing! Generate ANY OUTFITS!

Подписаться 150 тыс.

Просмотров 29 тыс.

50% 1

In this video, I'll show you how to train ANY Clothing and Outfits LORA SDXL for your Ai Characters using YOUR OWN IMAGES for Stable Diffusion! I spend hundreds of hours testing, experimenting, and hundreds of dollars in cloud computing training to bring you the ultimate clothes 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 any Clothes and Outfits you want with SDXL 1.0 LORA with the Kohya GUI tool.
What do you think of LORA Clothing training? Let me know in the comments!
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
SOCIAL MEDIA LINKS!
✨ Support my work on Patreon: / aitrepreneur
⚔️ Join the Discord server: bit.ly/aitdiscord
🧠 My Second Channel THE MAKER LAIR: bit.ly/themakerlair
📧 Business Contact: theaitrepreneur@gmail.com
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
✨ PATREON LINK: / aitrepreneur
Kohya ss GUI: github.com/bmaltais/kohya_ss
SDXL: huggingface.co/stabilityai/st...
Training Clothing presets: / clothes-outfits-90832934
Clothing list PDF: / 250-clothes-list-90833270
Exclusive LORA models: / 5-new-lora-90833826
Photoshop alternative: www.photopea.com/
Background remover: www.photoroom.com/tools/backg...
Previous LORA training video: • ULTIMATE SDXL LORA Tra...
Previous Style LORA training video: • SDXL LORA STYLE Traini...
scale_parameter=False relative_step=False warmup_init=False
Runpod: bit.ly/runpodAi
Runpod Template: runpod.io/gsc?template=ya6013...
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
►► My PC & Favorite Gear:
i9-12900K: amzn.to/3L03tLG
RTX 3090 Gigabyte Vision OC : amzn.to/40ANaue
SAMSUNG 980 PRO SSD 2TB PCIe NVMe: amzn.to/3oBR0WO
Kingston FURY Beast 64GB 3200MHz DDR4 : amzn.to/3osdZ6z
iCUE 4000X - White: amzn.to/40y9BAk
ASRock Z690 DDR4 : amzn.to/3Amcxph
Corsair RM850 - White : amzn.to/3NbXlm2
Corsair iCUE SP120 : amzn.to/43WR9nW
Noctua NH-D15 chromax.Black : amzn.to/3H7qQSa
EDUP PCIe WiFi 6E Card Bluetooth : amzn.to/40t5Lsk
Recording Gear:
Rode PodMic : amzn.to/43ZvYlm
Rode AI-1 USB Audio Interface : amzn.to/3N6ybFk
Rode WS2 Microphone Pop Filter : amzn.to/3oIo9Qw
Elgato Wave Mic Arm : amzn.to/3LosH7D
Stagg XLR Cable - Black - 6M : amzn.to/3L5Fuue
FetHead Microphone Preamp : amzn.to/41TWQ4o
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Special thanks to Royal Emperor:
- Totoro
- TNSEE
- RG
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:
►► • CHATGPT
RECOMMENDED WATCHING - My "Tutorial" Playlist:
►► bit.ly/TuTPlaylist
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Timestamps:
00:00:00 Train SDXL with LORA
00:00:53 What is LORA?
00:01:10 Koya SS GUI
00:01:53 Find the best images for LORA training
00:05:14 Image Dataset preparation
00:06:48 Remove the background of an image with Photoshop
00:09:45 Remove the background of an image without Photoshop for free
00:11:35 Increase the amount of training images by copying them
00:14:46 Upscale images sharper with Stable Diffusion
00:15:09 Avoid “zoomed in” image generation
00:15:47 SDXL LORA Training Folder Creation
00:16:42 Instance Prompt
00:18:31 Find the right clothes for LORA Training
00:19:30 LORA Repeats Training Steps
00:21:12 Regularization images
00:21:38 Folders Tab
00:22:11 Train Multiple items in one LORA
00:27:36 Image captioning for LORA training
00:32:13 How to replace a word in multiple text files on Windows
00:34:03 The right SDXL Clothes LORA Training Parameters
00:34:36 Train Batch size
00:35:24 Number of Epoch Training
00:35:30 BF16 & FP16 & Cache & LR Scheduler & Optimizer
00:35:55 Learning Rate for Clothing SDXL LORA
00:36:16 SDXL Image Resolution & Buckets
00:36:46 LORA Network Rank & Network alpha
00:37:28 Color augmentation & Flip Augmentation
00:39:06 LORA Training & VRAM Usage
00:39:38 GPU Solutions & RunPod
00:39:59 Select the right LORA Model
00:40:29 How to generate images of a custom ai character wearing custom ai clothes?
00:41:38 Conclusion on clothes LORA Training

Опубликовано:

6 июл 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 100

@Aitrepreneur 8 месяцев назад

HELLO HUMANS! Thank you for watching & do NOT forget to LIKE and SUBSCRIBE For More Ai Updates. Thx

@vitaliishebela2748 5 месяцев назад

Hey! Thank you for your work. Please tell me, is it possible to have exact details on clothing, line logos, and patterns?

@Xanderfied 8 месяцев назад

dunno why im watching this, my models never wear clothing

@irinaknyazeva4267 7 месяцев назад

😂😂😂

@LaidBackAI 7 месяцев назад

😅😅😅

@ArthurDNB 8 месяцев назад

Missed you. Good to see you are back

@Aitrepreneur 8 месяцев назад

Hey, thanks!

@sifu2u_now 8 месяцев назад

Wow...can't wait for the reveal.❤ Thank you Sir!

@SergeyTsyptsynPlus 8 месяцев назад

Absolutely unique content. Nothing can be even closer to the usefulness and deepness of you videos. Please continue making such a great videos.

@Aitrepreneur 8 месяцев назад

Thanks! Really appreciate it ;)

@mathieuwilson2985 8 месяцев назад

I cant believe this came out today. I am just finishing my cosplay studio and this was my objective. Wow ... I have so much clothes to train on and a mannequin... It will be insane. Thanks

@Aitrepreneur 8 месяцев назад

Glad I could help! ;)

@awais6044 8 месяцев назад

can you share a clothes items data share with me? i will be very grateful to you.

@graynell 8 месяцев назад

Exactly what i needed! Thanks!

@Aitrepreneur 8 месяцев назад

Glad it helped! :)

@romac9516 8 месяцев назад

My learning project is to create an accurate lora for Converse high-tops. I've wasted days doing the wrong things before learning from someone like you. So thanks very much for making the effort to make this. I'm much happier knowing it would still take experienced users a lot of time to prepare the training data.

@madrooky1398 7 месяцев назад

Spending a few days only to then learn from someone who took weeks is hardly a waste of time. I have spend moths to refine my process, most of the time with a broken dreambooth that simply did not work as expected. But it was never a waste of time because most of the prep work is on the data set, collecting and preparing high quality pictures with proper tagging and everything. On the way I learned so much more, Photoshop especially, which now also has the SD plugin to allow using SD models in PS. Just keep going and never think trying to learn is a waste of time. 💪

@chyldstudios 8 месяцев назад

Glad you came back. We missed our AI overlord!

@Aitrepreneur 8 месяцев назад

Glad to be back :)

@moshmoshpitpit 8 месяцев назад

Can’t think of AI image gen without thinking about your channel. Welcome back. You dropped this -> 👑

@Aitrepreneur 8 месяцев назад

Thanks! 😁

@76abbath 8 месяцев назад

I miss you! Happy to see you back!

@Aitrepreneur 8 месяцев назад

Thanks :)

@AIKnowledge2Go 8 месяцев назад

Hey K, your return brings a sigh of relief - breaks are vital but you had us worried. The depth and utility of your videos set a bar that's virtually unparalleled in the AI space. Keep leading us into the AI future, and remember: Your uniqueness is irreplaceable. 🚀

@Tj4ckinator 8 месяцев назад

Great video! Glad to see more great content Will this be the same workflow if you want to train a LoRa that creates a very specific characteristic like a specific kind of nose shape, teeth, one leg longer than the other, a thick neck or a body type or anything else very specific?

@varun_vashishtha 4 месяца назад

WHAT A VIDEO!!

@user-cw9ev9dq4v 6 месяцев назад

Thanks, that is what I searched

@elarcadenoah9000 8 месяцев назад

EXELENT VIDEO

@Aitrepreneur 8 месяцев назад

Thanks! ;)

@ImNotQualifiedToSayThisBut 8 месяцев назад

I too took a break, albeit from SD and not from youtube lol Glad to see you came back at the same time. Edit: Couldn't you also use Photoshop and do perspective manipulation to add more images to your dataset?

@user-yo3cp8ju4y 8 месяцев назад

@Aitrepreneur I'm planning to make a LoRA for the 4 Keijo school swimsuits. Should I use the swimsuits for the anime or cosplay clothes? Or is it better if I use both? Also, are you planning to make a video about SDXL Checkpoint training?

@johnny1tap 7 месяцев назад

you are so awesome.

@jeethechamp7458 8 месяцев назад

Thanks

@lilillllii246 6 месяцев назад

I created a clothing lora by referring to your video, but when I applied inpaint to apply it to the model in the existing image, the clothes did not fit the clothing roller. What should I do?

@diggydretv 5 месяцев назад

I have adjusted the learning rate between 0.0006 and 0.0009. I have changed my network rank between 64 and 128. For some reason, no matter what I do, I get blurry backgrounds when using the LoRA. The backgrounds don't start to become clear until I get to an epoch where the model is showing the early signs of burning. What am I doing wrong?

@aimademerich 8 месяцев назад

You are 👑 😂 Thank you

@ysy69 8 месяцев назад

Thank you for your yet again comprehensive tutorial. Regarding the Train Batch Size, would it be the same for Human subjects? Higher the TBS, more flexible but less likeness ?

@Aitrepreneur 8 месяцев назад

more or less yes, the best batch size would be 2 also but 1 is almost as good. However for style training, batch size of 5 is still the best, I haven't seen any big difference in quality or flexibility, and since it makes the training so much more faster, better leave it at 5 for style

@AnnCatsanndra 7 месяцев назад

What Instance/Class prompts should I use for training an action / behavior / concept?

@odev6764 6 месяцев назад

I really enjoyed the idea. Is it possible to finetune a lora to work with a controlnet and openpose? the ideia is: I select a person and want to keep the same person, background and pose, but only change the clothes. I is more like a face swap but for clothes and of course it needs to adjust by the person body. like a dressing room

@karkarkarkarkar 8 месяцев назад

Can you do a tutorial on training LoRAs that don't have pre-existing concepts inside SDXL? I'm trying to train a lora for "four armed humanoid" and I can't get it to unlearn its good anatomy.

@user-dx8qi6hp6i 3 месяца назад

Is it possible to "Exact" swap clothes by uploading clothes picture and model picture (both as input pic)?

@thanksfernuthin 8 месяцев назад

I trained two characters as a test with your first video for that. I've been playing with them ever since. The results were amazing and I'd like to really do a great job and pick the best training images... I'll get to it eventually. My question is... is the vast improvement in training due to SDXL? Or will SD1.5 train just as well with the latest Kohya and your methods?

@Aitrepreneur 8 месяцев назад

SDXL is really amazing yes, much more powerful and flexible to train than 1.5 but the principle is the same for 1.5 also, the only difference might be some of the training parameters such as the learning rate that should be a bit lower (try 0.0001) and the resolution of 512x512 instead of 1024x1024

@thanksfernuthin 8 месяцев назад

@@Aitrepreneur Thanks K.

@RonnieMirands 8 месяцев назад

Wow, thanks for your amazing video. You said the minimun is 10. I am trying to train complex cloth of a carmelite nun, that has 4 different parts, a tunic, a coif, a veil and a mantle. But not getting good results yet. The last training i tried i used 47 images! Maybe i need more like 100 images? :(

@RoyJNg 7 месяцев назад

Hi, is the same concept here can be applied to making mask on faces like scuba mask, goggles, or masquerade mask?

@NeogreenPeaches 8 месяцев назад

It was to my understanding that the training image folder names don't affect the actual instance/class prompts if you include caption files. The only thing that matters if you have caption files is the "_" prefix. In this case you would need to include the prompts at the very beginning of your caption file. Am I wrong in this assumption? There is so much misinformation floating around that it is near impossible to get a straight answer on this.

@DannyCaruzo 2 месяца назад

I want to make a clothing lora too,but for SD 1.5, is the process and the settings the same ? I'm very thankful about any help ❤ thank you for the great video

@awais6044 Месяц назад

You find answer?

@vilainsinge5282 5 месяцев назад

I want to train a LORA based on a face mask that i've crafted. I dont know if i should train it a a person - with me wearing it, or as an object - to use it on different people. If i train it as an object, i considered put the reference images of me wearing it and specify that it's worn by someone in the description. Any advices ?

@shashanksrivastava8638 5 месяцев назад

how can we download the .ckpt file or the .safetensor file for the trained model??

@daviddelpinoherrera2672 6 месяцев назад

I have a 3D model of a cyberpunk style clothes made in Marvelous Designer for Blender that I want to render to use as dataset for this. It looks realistic, but I want to use it in anime models. Is that possible?

@peromiestiloesunico 3 месяца назад

Does this can work with stable diffusion 2.1 and 1.8

@matthewanacleto7885 Месяц назад

Will this work for ropes and other restraints?

@fedmartinez4077 8 месяцев назад

tried creating a Hoodie LoRA by following this method for practice, but when i finally tested it, the resulting images leave out the face just like the ones in my dataset almost all the time. i added "out of frame", "faceless", in the negative prompt and added weights also to it, but it's still generating the attire more than a person wearing it. any tips?

@rahulsundar7171 6 месяцев назад

Model is overtrained

@nefwaenre 8 месяцев назад

Thank you for this!! Will you also be doing clothing training for non SDXL LoRA? i *really need* some clothing for my dudes. One of my dudes wear shirts with cute patterns, so what i do is, at first manually stamp the pattern on his shirt and then run it through SD. But say, when the original picture has no shirt/tank tops, i have no idea on how to put on shirts on the guy. Any help on how to do this? 🥺 And finally i'd be able to train some Hogwarts robes and Quidditch outfits for my dudes! 🥺

@Aitrepreneur 8 месяцев назад

The principle is the exact same, the only difference might be the learning rate (try something like 0.0001) and the resolution of 512x512 instead of 1024x1024

@nefwaenre 8 месяцев назад

@@Aitrepreneur Thank you!!

@chenchangye5467 8 месяцев назад

what if i traning with dataset of 50 to 100 img? the preset json and lowering repeat doesnt seem work

@Aitrepreneur 8 месяцев назад

the number of images is not an issue, the principle is always the same

@dthSinthoras 8 месяцев назад

Would hairstyles work the same way?

@harveythe8053 8 месяцев назад

How would you go about doing this for an entire set of armor? Or soldiers outfits, or say Space Marine armor, big bulky metal armor? Would you have to do it piece by piece and a different lora for each section, or will images wearing an entire armor set work?

@Aitrepreneur 8 месяцев назад

just like the peaky blinder outfit, full body armor images + zoomed in images + another folder with the helmet

@harveythe8053 8 месяцев назад

Appreciate the reply! I'll give it a go wit that method :) Thanks a bunch@@Aitrepreneur

@horrid13 8 месяцев назад

will this work for tattoos?

@AnandhuAK 3 месяца назад

How train a pose?

@sleepyyfoxx7041 8 месяцев назад

How would I go about using a character, style, and clothing LoRA in one generation? Would I need to train a LoRA on all 3 of these or is there a way to combine LoRA models?

@deadlymarmoset2074 8 месяцев назад

There is a "super merger" extension for automatic1111 that can combine loras, however, you can just use multiple loras in the prompt at the same time.

@sleepyyfoxx7041 8 месяцев назад

@@deadlymarmoset2074 Thanks for the info!!

@leavemealoneandgoaway 8 месяцев назад

so have you tried getting chatgpt to write the descriptions for each image and training with those? might shortcut the worst part about training.

@Aitrepreneur 8 месяцев назад

yes but chatgpt vison is still not available to everyone (just got mine)

@quercus3290 8 месяцев назад

personaly I want my training images with as much details as possible if you dont set your network dimension too high and have a wide variety of images with proper tagging, ive never had issue with SD confusing background elements as part of any clothing.

@Aitrepreneur 8 месяцев назад

It actually often happens in those 2 scenarios: 1) The background is often the same, same settings, same lighting, etc so it will get picked on during training and will often be present each time you try generating an image 2) Sometimes the instance prompt you chose can influence the settings also, for example if instead of choosing for the yellow cyberpunk jacket I put "cyberpunk 2077 jacket" you would often generate images of cyberpunk cities in the background you technically don't need to remove the background if the variety is there, but I found it's safer this way

@quercus3290 8 месяцев назад

@@Aitrepreneur got any loras on civit mate?

@CalvinSturm 8 месяцев назад

Why have you been slowing down on uploads? :(

@jwc8963 8 месяцев назад

Are those training parameters the same if I use SD1.5 base model for training? I have some problems when using SD v1.6 and try to avoid using XL models🤦🏻‍♂️

@Aitrepreneur 8 месяцев назад

the learning rate should probably be a little lower, so something like 0.0001, and the resolution needs to be 512x512, other than that yes I think it should work

@jwc8963 8 месяцев назад

@@Aitrepreneur the picture in dataset can be in different resolution when training based on XL model?

@lefourbe5596 8 месяцев назад

@@jwc8963 you can just leave the default setting taking car of downscaling your pictures so don't worry. it's better than birme

@jwc8963 8 месяцев назад

@@Aitrepreneur should I increase the epoch number as a smaller LR is used?

@Aitrepreneur 8 месяцев назад

no need 15 is already way more than enough

@3k3k3 8 месяцев назад

Where did you go...???? So many questions..

@Aitrepreneur 8 месяцев назад

Just working on the video :)

@HowToRealise 8 месяцев назад

Why do you cover faces with black? Why not white/average/blurred/grey?

@Aitrepreneur 8 месяцев назад

doesn't matter what you use as long as the face is not visible at all in the image :)

@ronbere 8 месяцев назад

kohya stopped working all of a sudden, I hope to find an alternative to this buggy program

@Aitrepreneur 8 месяцев назад

What error you got?

@ronbere 8 месяцев назад

@@Aitrepreneur returned non-zero exit status 1… I've reinstalled it several times on different disks, and the error is still there.

@Aitrepreneur 8 месяцев назад

send me a dm on discord

@ronbere 8 месяцев назад

@@Aitrepreneur done

@eddie9965 8 месяцев назад

First

@vokuh 8 месяцев назад

@patrickderp1044 8 месяцев назад

the peaky blinders suit is cringe. the cyberpunk clothes are cringe. but the results are undisputable, we were all wondering if you died but of course you were doing a month of training, and now hopefully it takes us less than a month to replicate this because of the video. thank you!

@Aitrepreneur 8 месяцев назад

How dare you! You know I'm cringe also so training cringe clothes is just natural for me 😎