One Click SDXL LoRA Training is HERE! Easiest Method
HTML-код
- Опубликовано: 7 авг 2024
- 🎨 PixelDojo: www.pixeldojo.ai/?...
Join the discord! / discord
Training a LoRA model for SDXL can often be a lengthy and challenging task, but I've developed a game-changing tool that simplifies this process immensely. With just a single click, you can now generate high-quality models for styles, objects, and people. Join me in this comprehensive tutorial where I walk you through the entire process from start to finish. Whether you're a beginner or an experienced user, this video will provide valuable insights and tips to make your model training efficient and effective. Don't miss out on discovering how to elevate your SDXL projects with ease!
🚀 Highlights of this video:
Introduction to the new tool and its capabilities
Step-by-step guide to generating high-quality models
Tips and tricks for optimizing your workflow
Real-time demonstration of the tool in action
Solutions to common challenges in training LoRA models
🔔 Subscribe to my channel for more tutorials, tips, and updates on AI.
👍 If you find this video helpful, don't forget to like, share, and leave a comment with your thoughts or questions. Your feedback helps me create better content for you!
#LoRAModels #SDXL #3DPrinting #AI #HomeAutomation #SelfDrivingCars #ArtificialIntelligence #Tutorial
Image enhanced in the video with the face enhancer: pixeldojo.ai/community-galler... 
Наука
There are instructions on the Style Trainer page already covering this, but in general, if ANY of your images are low resolution (less than 1024 width or 1024 height), you will need to uncheck the face cropping box, otherwise you will get an error. Stable diffusion XL expects images that are at least 1024X1024. Ideally, you would just use better images lol. Low quality in = low quality out.
link to the style/LoRA trainer: pixeldojo.ai/lora-trainer
links to News which covers how to train a LoRA for an object in 7 minutes: pixeldojo.ai/news
FAQ:
Q: Which model does this run on?
A: SDXL
Q: Can I train on other base models?
A: Let me know what you are looking for. I add new features almost daily.
Q: I get an error on upload!
A: Make sure you are uploading a ZIP file as the form states. ZIP should be less than 100MB. All images need to be 1024X1024 or larger otherwise uncheck the face crop feature (also try to find higher quality images)
Q: Is this free?
A: No, it's a subscription as covered in all of my other videos. GPU's are expensive. This is the lowest cost suite of AI tools on the market.
Q: Is there a free trial?
A: No, GPU's are expensive. This is the lowest cost suite of AI tools on the market
Amazing! I'll try this today or latest tomorrow 🥰 thank you so much!
Finally! Someone did it! I can finally breathe. Thank you so much for creating this tool - I've been waiting for it for ages. I have so much love and respect for your work. I'm definitely going to try it out, and after watching your video, I'm already 80% confident it's going to be amazing!
Definitely let me know what you think. I'm constantly adding new features and improving :)
thanks for sharing this wonderfull tool
Can't wait to try this. Trying to keep up. Another winner! 👍
What if I do not want my pictures to be cropped (when choosing Person). What I want a full body Lora ? Also, can I download the loras/photos?
Great job on this
Thank you
This is amazing! nice job!
Thank you!!
Thank you! 🙂
Very cool features! As I can see you as an expert have an eye for the nifty details of the workflow and provided an solution for it! Really nice!
Thank you for the kind words :)
Fantastic Thank you. Is there a change to have this training for sd 1.5 Models?
I'll see what I can do :)
good stuff!
Amazing:) Do you have a way to see the prompt variables that are trained into the lora? Also could you enable manual manipulation of the captioning incase we only want certain parts of an image to be used.
Let me see what I can do!
hey sir does the training of lora in website is free? or i need to buy subscription to complete the training cause when i start training it says "something went wrong"..
Yes, you must be subscribed to use the trainer. Nvidia A100’s are expensive haha
You the GOAT this is exactly what i wanted 😊
Awesome!
Is there a way to download the generated Lora ? so I can use them inside comfyui. Does it work with anime style, for example if I screenshot some anime and upload them onto this "lora maker" !? will it support ponyXL by the way ?
You can definitely download and use the LoRA. I have not tested ponyXL. It trains on top of stable diffusion xl currently.
I subscribed the dojopixel, but when I downloaded the lora and put it with my sambhalXL model, it is not giving proper result, which model i should use with the generated lora
Most models based on SDXL base should work. I would start with SDXL and experiment from there
@@allyourtechai i have used sd_xl_base_1.0, shambhalaxl, dreamshaperXL10_alpha2xl10, all three model are not working at all, however your website is giving amazing results with the same lora, now please tell us which model you are using so that my result are as goos as the result in your website
@@allyourtechai i have used sd_xl_base_1.0, shambhalaxl, dreamshaperXL10_alpha2xl10, all three model are not working at all, however your website is giving amazing results with the same lora, now please tell us which model you are using so that my result are as goos as the result in your website
Hi I have subscribed and the website is giving good results after I trained my lora, can you please tell me exactly which sdxl lora you're using. because we have tried with a few sdxl models but the results are horrific on stable diffusion. Please help figure out which exactly sdxl model you use to get such amazing photos on your website, thank you
Using SDXL base actually
@@allyourtechai Hi so I used SDXL base 1.0, the results are not showing at all, what version of sdxl should I use to get results like them in the website
@@chanbose9674 I've had this issue too. In my case it wasn't he base model, it were the settings I've used. The optimizer and precision to be precise. The results were very bad at first. I've then changed them close to the ones the creator of the video explained in another video and it gave me much better results. I don't have my PC rn, but I think I've used adafactor and fp16.
Thanks! Do you have to be subscribed to create a LORA? I don't mind paying the $25 but I want to know it works well before I pay.
You do. The GPU time is fairly expensive, so i'm not able to offer free credits without raising the price which I would like to avoid if possible. YOu can always join the discord and talk to customers if you would like. I have a couple hundred happy subscribers since launch a few weeks ago. I also have the community gallery on the site with thousands of images created by the community if you would like to check those out as well.
@@allyourtechai thanks, I’ll do that
Btw i noticed that Stable Diffusion changes the eyebrows and likes to exaggerate and arch them too much sometimes. It's noticeable on the generated images at 9:47, notice the difference compared to the original girl. I also get this with stable diffusion when I run control nets that replace face.
Does this work with product training? Also, considering that details are key for e-commerce, how much hallucination do you get on these Loras after they’re trained?
I haven’t tested that scenario yet. Maybe you can be the first :)
My latest video shows a style training with a car. You can see the process and modify for products im sure
@@bloveaffiliateI tried this method before and got similar results. The problem is, you can’t sell something that looks like a product… it’s all about the details of the real product. I’m sure soon there’s gonna be a way to train a model without any AI hallucinations. But if you check the latest video I uploaded, you can see my workflow currently. Which uses AI and traditional tools together.
I am going to try to work with kohya for 7 days. I hope this video will help me. My gpu is 2060 super.
Sd 1.5 is what I recommend
what if you want to use another base model? what even is the current base model used?
As the title of the video states, SDXL. I'm adding other base models as well. Likely starting with realvisXL. Join the discord to vote on a model if you would like.
Well the upscale at 10:44 seems to fix the ugly exaggerated eyebrow arch, now it's much better 😁👏👏👏 hope my comments will signal to YT algorithm the importance of this video 😁👌
Upscaler makes a huge difference
Is stable diffusion still the best offline image generator? What I mean by offline is the ability to install it locally.
From what I have seen, yes. There are some new models with different underlying technology, but I haven’t seen the image quality match or beat a stable diffusion fine tune
@@allyourtechai Thank you for the info!
This looks really great, could you add support for creating Lora's for Pony models ?
You can create style or object LoRAs, so you should be covered
I keep getting an error when I start training. Guessing there is a max amount of photos or zip file size?? Thanks!! (UPDATE: had something to do with the "Use Face Detection" box being checked.)
If your base image is less than 1024 pixels the face detection can’t crop it down and throws an error. Glad you got it resolved. I’ll add a note about it to the ui
@@allyourtechai Just wanted you to know...under Style Image Creator, My Custom Styles... in the "Click on" sentence, apply is spelled wrong 😉. Loving Pixeldojo so far bud!!!
Start training doesn't come up on my screen, there's only a greyed out "subscribe to X".
Once you subscribe you can use all of the tools
I think it's fair you charge for this, but would be worth mentioning in the video, perhaps even let us have one free shot at it. Went through the entire process only to be offered a "Subscribed to train" button that doesn't do anything. Make it clear as the TOP of the page, at least. Another thing, what data is used for the regulization data when creating a style?
Great feedback, thank you
Bug report: it worked once, got an LoRA downloaded. I wanted to try again with different images, but it stopped working. The "Start Training" no longer started the training, no UI feedback, or error reported.
Was it a large file upload? I need to add an “uploading” indicator. It currently doesn’t show feedback while the zip is uploading which isn’t ideal. Are you on the discord?
discord.gg/sBgBuhgc4c
Please bring inpaint and outpaint feature to pixel dozo !!
Working on it!
I wish we had a trial or at least ONE free per day or something, call me cheap but I will not blindly pay 25$usd (over 35$ in my currency) upfront to test it.
Until then I will keep training locally with what I know will work though, albeit more tedious and time consuming.
Great vid and congratulations if your product is as good as advertised :)
It costs around $9 to train a LoRA on an Nvidia H100 GPU, so no chance of a free trial. I invite you to join the discord and talk to real customers though.
@@allyourtechai But subs are 25/month? you are banking on people doing less and 3 loras PER MONTH to make money or break even or do you also have another money gate behind the subscription to account for that? I'm even more scare to sub now!
Edit, Also noticed on the page you can't choose the model used? is there a reason for that? I'd love to use my custom model merges/models i download
PS :Not trying to be a dick here, it's very hard with youtube comments but I swear i'm just asking and not being passive aggressive lol
Unlimited or one time only?
Unlimited
@@allyourtechai unlimited not working.only one time work
@@WorldCuppppp582 You only have unlimited generations with a subscription.
Please more videos
More coming!
@@allyourtechai thanks 🙏
Is lora training unlimited?
Yes
There is no Lora training option available any more additionally it do not even create one picture without subscription
This isn’t true. Click on training in the menu.
Correct that there is and has always been a subscription. It run on $25,000 GPUs that don’t pay for themselves :)
@@allyourtechai I do not know if it is region specified or something like that, I can only see Style Trainer and Style image Creator in "Train" Tab I am from Pakistan
why not just github their script. Not handing over training data to some mystery third party.
I hope, my friend, that you will try the free site once and return this free paragraph to the site
?
@@allyourtechai Your site had a one-time use for free to edit the image now a subscription is required in order to use the tools I hope to return a one-time use of the site ( Free)
@@allyourtechai Your site had a one-time use for free to edit the image now a subscription is required in order to use the tools I hope to return a one-time use of the site ( Free)
Your site had a one-time use for free to edit the image now a subscription is required in order to use the tools I hope to return a one-time use of the site ( Free)
@@allyourtechaiYour site had a one-time use for free to edit the image now a subscription is required in order to use the tools I hope to return a one-time use of the site ( Free)
is this free?
The subscription is $25 per month for unlimited use of all of the tools
Charges a subscription fee and having difficulties with the Lora Training.
Over 300 LoRA’s trained just in the last 4 days. What issue did you have? I did see in the logs that someone was trying to upload images even though it says in 3 places to upload a zip file
@@allyourtechai It keeps crashing at 2 minutes on the analzying. I set up ten anime images of a character and it freezes and closes. I tried to create a style and it does the same. Any restrictions on images? Do I need to crop them to certain sizes?
You should join the discord. Hard to reply back on RUclips, but generally your images need to be over 1024x1024 IF you enable face crop. If you disable the cropping size doesn’t matter, but low quality input will result in low quality output. If your images are small use the upscaler first if you can.
@@allyourtechaiI shall do that later today hopefully will be able to figure out my issue 😂
@@kevinpeake Sounds good, let me know if you run into issues. This is the only problem I have seen anyone run into (low resolution input images). So you should be good once you check that out.
Can’t believe the entitlement in these comments.
"Hey guys I've devoted the past month to this!"
"WHY ISN'T IT BETTER THAN EVERYTHING?!"
"WHY DO YOU WANT MY MONEYS?!"
Thank you for your hard work Sir.
Ignore the angry children.
didnt work at all 10 pictures failed
You have to have a subscription in order to generate images
Doesn't work. Tried training a lora 3 times, its kept failing. unsubbed from pixeldojo
You should have reached out to me. It's super easy and more than likely a simple mistake with your images. Train with 8-10 high quality images and it works every time. There have been over 900 LoRA's trained in the last 30 days now.
Great job! But $25 monthly after just 1 time try image generation is greedy 🙄
I spend thousands per month on GPU's for the service. Last month I finally broke even. If that's greedy, i'm doing an awful job at it. It's 15 different tools in one service and less expensive than any of the basic upscalers on the market.
Its not greedy. With the speed this enables, without hassle? Its perfectly fair.
@@allyourtechai Just an idea: perhaps you could fine-tune the pricing and add a limited tier for a reduced fee...
Right, and very soon, this will ask me for a subscription… after it gets nerfed so much it becomes useless… I’ll stick with local use so I have more control.
Nerfed? I’m the creator of the software. I work several hours a day adding new features. Check out the discord to see what customers think.
Feel free to run everything locally as well. That’s why I keep creating new local guides as well. My goal is to make these tools accessible to everyone including those without a GPU at home.
Not very photorealistic. Needs a lot of optimization
Input images matter as well as prompts. Best part though is that the creative upscaler can be used to really make the images look great
Paid service, no ability to tag each image separately, seemingly can't even choose what model to train over. No thanks! I'll continue to use the local software that this video describes as too difficult to learn.
...and I do turotials on every alternative method of training a LoRA. I get over 100 comments a week from people that have issues, which is what prompted me to create this setup. If you are able to run everything locally and you are happy with it, awesome! I'll continue to post great tutorials for local training.
Which models would you have liked to train over on PixelDojo? If you check the discord I add new features almost daily, so the service is always improving and staying ahead of the rest. If you have feedback or a specific feature request, let me know. even if you don't use it.
only makes videos now shilling his site lol buy an ad
I've been up coding all night every night for the past 4 months working on this stuff because I was frustrated by the lack of accessibility of these tools. So yeah, i'm excited to let people know about the problems i'm solving.
@@allyourtechaiIt's totally fair to show all the cool feature with all your passion. And it's really cool! But you should drop a line in the video an description about the conditions and pricing to be fair. If one has to subscribe first (there isn't even a pricing tab before) that's not fair...
Home page has an entire pricing section. Thanks for the feedback, I’ll be sure to make it more clear in the future. I’m just trying to cover my server costs.
Wow 🤯 Thanks for sharing 👊🏻🫡
Any time!