#### Link from my Video #### Join my Discord: discord.gg/XKAk7GUzAW Bulk Resize: bulkresizephotos.com/en Kohya-ss Install: github.com/bmaltais/kohya_ss Booru Dataset Tag Manager: github.com/starik222/BooruDatasetTagManager/releases/tag/v1.6.3 SD 1.5 Training Model: huggingface.co/runwayml/stable-diffusion-v1-5/tree/main
Great video, but aren't HyperNetwork's also an option? Is there any reason why we'd use Lora's over Hyper Networks? So far I've gotten better results training HyperNetworks, but admittedly Lora's had only just arrived on the scene when I was experimenting.
do you know how feasible it is to install Stable Diffusion on a Mac? for example Mac Mini M2 Pro? Will it work by running it on a virtual machine to windows or will it be better via bootcamp? or is there a Mac direct version?
All your videos, and especially this, are invaluable to the AI community. Foremost the beginners, but also the enthusiasts that are kept up to date with the rapidly changing and improving technology. And explicitly worth mentioning: Love your enunciation! Your well-phrased, concise use of language and well prepared scripting makes your videos efficient and a pleasure to watch! Thank you so much for your efforts and contribution!
A really helpful and informative video. In addition, and I find this particularly important, your pronunciation is also really super. For me, as a non-native speaker, this is very important. In many videos, people just mumble to themselves, or you can not understand them so well for technical reasons. With your videos I found the first time someone I can really understand excellent and crystal clear. 😊
the taging changes a fair bit depending on if your trying to train a topic or a style or a specific person and it would be nice to show the differences there. i see in on another video they explained how you should tag everything you DONT want to be a part of your lora unless you are training for a single person or thing, and I wold like to see a more in depth dive into tagging and its effects
The tagging strategy is a pretty confusing. I've also seen tutorials that say you should tag only the things you DO want to be part of your lora/model. And some tutorials say to not use tags at all and let the AI figure it all out. And should you use a unique "keyword" tag to train it to that particular unique "keyword" to use in the prompts? I've tried different the approaches but didn't get any conclusive results. They were all pretty bad and didn't work as expected. 🤨 A video from Olivio covering just the tagging aspect with a demo for each approach would be awesome.
@@Beltalowda55 captions are what a model uses to try and predict what the picture should look like during training. Difference between predicted and actual image is used to adjust the model in the right direction. So it should be obvious that you want to describe everything that is not a part of the concept you're trying to learn, to help the model to make a correct prediction and narrow the direction for the learning. Keyword is added for the model to associate the concept with that keyword, that helps to preserve sanity. It sould be noted though that the training process does not actually differentiate what you're trying to learn in any way. All it does is it tries to match a predicted image with the actual image. That means it can associate to your concept any other or multiple tags you used in unpredictable ways, basically splitting the concept between them. To mitigate that, no tag except a keyword should be used too frequently (in my experience 30-50% is the limit)
@@fritt_wastakenI wish you had some videos up to talk about this. I think I understand what you're saying but some of it is a little fuzzy. Thank you for the info nonetheless.
Te trouble with all these tutorials is that when I spend an entire day following them, the software I see does not ever seem to match the software in the video. When I reach an error screen, I never know why and I have no comprehension of what to do about it. It's quite daunting for an artist who thought the only thing he would have to master in life was using a pencil and a brush. :(
@@ScaryStoriesNYC I absolutly understand that. Software is developing so fast and you have to make tutorial more modular, to change the parts, whenever something changes in Software.
Agreed, he gets the pace just right. Not easy to do, some people take too much time with no benefit, some too fast and hard to follow. The balance here is just right 👌
24:08 the same when you said "512x512" is in the past, the same goes to folder structure, is not necessary to create 3 folders, only the image folder and the inside folder with your dataset, for model and log you can leave the main folder "Neon" and the files will be saved there with no problems, I discovered that by accident and no issues after.
You are my hero today! ❤ I was searching for exact this topic weekly for last 2 months 😂 Last time was this morning. Not sure if I go to sleep now or give it a try 😅
Nice VDO as usual. You mentioned in the VDO that you will cover both Dreambooth & Dreamboot LoRA; however, there has been no Dreamboot LoRA covered. I am not sure whether both Dreambooth & Dreamboot LoRA will have the same process & pretrain model or not? For art painting style training using Dreamboot LoRA, could you please suggest an appropriate Pretrained Model, and min. total step for each Epoch. I tried 100 Steps for sample 20 images (2000 in total) but when I tested with text prompts used in the training, the results were still far from original images. Also tried with LyCORIS/LoCon but with no success. I would be very appreciated if you could demonstrate how to train a painting style like Van Gogh painting style (though it is already available in some models) showing that the result from prompts used in the training process can reproduce images used in the training.
Dreambooth doesn't make LORAs. It trains a model, then you can extract a LORA from that by comparing the original base model and the trained model. You can see the tabs for that in the video when he's looking over the Dreambooth GUI.
Danke! :) Thanks Olivio for your many great explanation videos and tutorials and that you made the effort to really go through every step of the setup and first usage of kohya. I needed exactly this to get to a starting point of the mysterious journey of training my own models/loras and now I can finally experiment on my own.
@@plattepus Try installing and running SD, and watch for the error in the console, It will give you a pretty noticeable verbose error saying Pip is not installed. If so just run the python install again and make sure to tick the box to install pip in the menu. Hope that helps.
@@addermoth appreciate it man. Also, I'm only seeing "gui.bat" in the "kohya_ss" folder, and not "user gui.bat", but it still seems to launch it fine. And I can't tell if it installed cudnn, I put the cuddn_windows into kohya_ss, ran the command to install it, but it never said "done" or anything.
Good evening Olivio, So I just followed your guide to create my first LORA but I find myself with 22 hours of training! I must have made a mistake somewhere ;-) However, I have an RTX 4070 TI and a Core I9 13900k. 120 photos of the same model in high quality (FHD) half are faces and the other is the full length model. I chose 10 steps, I think my concern comes from there. I'm still going to let it finish just out of curiosity. Thank you for the guide which will have nevertheless been very useful to me.
The Kohya page and download isn't the same as in this video and the setup.bat file has been updated. It doesn't ask me the same questions as it does him in the video. How did you get it to work?
Correct, but don't use convert RGBA To RGB and transparent image to white Background, because it's make ur trigger background situation or etc, at your prompt be blank or just white or just template take a sample data.
This is awesome, thanks. I think you did a great job on disseminating the critical information concisely, I wish more tutorials were done like this. I have a question, how would training for an illustration style, using real photos for face reference differ, if at all?
Thanks for the video Olivio! I got a runtime error when trying to train a lora and this is how i (think) i fixed it. When installing kohya, Olivio said that the questions prompts are correctly pre selected. This wasnt the case for me. The question prompt that has a selection for fp16 had to be selected during installation. It seemed that otherwise kohya was trying to use my CPU instead of my GPU. To fix my error, i deleted kohya install folder, reinstalled kohya, uninstalled and reinstalled torch when prompted, and selected fp16 when prompted. i started training a lora before writing this post and its still working so if you have a runtime error like i did, try this maybe. hope this helps!
By far the most thorough and easiest to understand guide. I followed someone else's this past week and they skipped. The Cudda setup and explanation The Folder Structuring setup explanation The explanation of the different tabs, up until now I thought the Dreambooth was just a different way to train Lora's....anyway their guide was really confusing and missed a lot of stuff you cover.
If you are like me and have a hard time finding high res photos of the target person you are trying to train, simply run all the images through face restoration. It will add all the detail you need for training. Hope this helps someone.
@@gohan2091 Topaz Gigapixel, or more like all in one tool: Topaz Photo AI. It has upscaler, sharpener, denoiser and face restoration all together. (not getting payed for this, i simply love to use this by myself)
i would suggest to use ligthroom to create keywords and export with filenames containing comma-separated keywords, and create captioning .txt files from that filenames with a simple python script 👍🏻 Much better keywording and organization. (smart albums, etc) Thanks
Fantastic tutorial Olivio! Love all your AI videos. Can you use LORA to train other objects and keep them consistent? Using your technique, I've been trying to create a LORA of a specific baseball glove and a specific helmet, and apply it to an image of a baseball player in stadium. Most of the outputs I get are quite monstrous like 10 finger alien gloves. I've given it a ton of data like close up/medium shots of the glove from multiple angles, medium and wide shot of someone wearing the glove. Still nothing works. Any thoughts on how to train a LORA to keep something like a glove consistent? I'm using using SD v1.5 pruned, photon v1, realistic vision v40 as the checkpoints in A1111.
Hello friend, is there going to be anything new on Lora training program for people? I haven't heard anything in a long time, maybe there are some new software chips
Great video Olivio! As I am watching this, I wonder what goes into the "log" folder and "model" folder? I didn't see what those are for. Also, do we need regularization images? If so, how many?
24:15 When Olli said "The images are not in the images folder" I got really confused. Eventually I figured out that you still have to copy your source images into the new folder, otherwise it won't have anything to tag.
Great video! But I prefer to train Textual Inversion for characters (a specific person) in Automatic1111: it's way faster, simpler, I get the same result than a Lora and spend less HDD space and only 2 tokens for the prompt!
Is the step of "merge" the models also necessary with Lora? I got confused because I thought you were going to create a Lora but you ended up creating a Checkpoint.
Hey Olivio, great vidio. Really go deep and surface all in one. I have a question. The style model I want to use is Dreamlike Photorealistic, from the same people as realisticvision. In your case you trained with sd15 in Kohya-Dreambooth (with captions for training dataset), but your video is entitled Lora. With Lora I usually use class dataset and Training dataset. Not sure if that was intentional. But more to the point, should i train with SD15 and merge to Dreamlike, go apply that style? Should I also use captions? I am doing Lora,thought. Thanks
If I understand correctly, the merge is just a weighted average of the model weights, so the shapes of the models must be the same. So you should merge your LoRA with another LoRA that fits the style you like
Hi, thanks for your amazing videos. But I had a lot of problems. first, the setup: you didnt show us all the ssettings in the cmd, i was really confused, second the picture tagging: it just spit out some errors. and I am really stuck at this point. So I couldnt train any modells :(
Perfect tutorial. Thanks for your work 😊🙏 I have noticed there is now also LoHa, have you tested, compared? Otherwise I’m very happy to use this workflow
The problem with Kohya is that is updating often and the guy like to move around and change things in installation, interface, parameters how to name things. This is the 5th tutorial I try to follow... all of them have something different from the Kohya I have installed yesterday. But all of them were usefull to puzzle out how to make it... storta work. I was able to make the first lora with bad results, but probably as I'm not able to load the reference pictures for some reason. It gives an error and do not use them.
Great, but you skipped some confusing stuff....do I choose no distributed choice with a 4090 card, do I choose to optimize the script with dynomo, and which dynamo if so? Do I use DEEPSPEED, etc...I have no idea about these cryptic choices and what would be optimal... 😞 help?
Hey Legend! I fell in love with your tutorials! I have a question: this is not working on my laptop and on my server as well :( I have 2080 RTX in both. Can this be the reason? In my server there are 2x2070... do you have any other method to train models for stable diffusion?
Hi Olivio, thanks for this guide, it's a bundle of usefull informations. I have a question: at 30:56, when there's your Kohya full screen, I read "Learning rate: 0.00001", the default should be 0.0001, is this correct?
Thanks. This video helped a lot. However i do have aproblem with Kohya. All the buttons are not working so cannot really do anything. The is a warnig about LD Library in Terminbal. Am i missing some files or what is going on?
I don't understand why an app can give better results with 10 images that are very similair but SD makes me look like a totally different person. even the tiktok filter gives better results with only 1 image (granted, it changes your look a little too much) but still gives a result where you at least look human. SD is much harder to train than I expected.
Thank you for the video! You mentioned you could do one with an online server, please, pleeeeeeease would you be so kind as to show us how to train a LORA step by step with RunPod ? 😭😭😭
If I'm trying to train on an anime character should I use background removal to isolate her? I have many illustrations of the character with design elements around her rather than full color backgrounds and I'd like to use her LORA to do full color art and more. Thanks for the great guide!
#### Link from my Video ####
Join my Discord: discord.gg/XKAk7GUzAW
Bulk Resize: bulkresizephotos.com/en
Kohya-ss Install: github.com/bmaltais/kohya_ss
Booru Dataset Tag Manager: github.com/starik222/BooruDatasetTagManager/releases/tag/v1.6.3
SD 1.5 Training Model: huggingface.co/runwayml/stable-diffusion-v1-5/tree/main
make tutorial finetune on kohya please..
Great video, but aren't HyperNetwork's also an option? Is there any reason why we'd use Lora's over Hyper Networks? So far I've gotten better results training HyperNetworks, but admittedly Lora's had only just arrived on the scene when I was experimenting.
everytime I paste the cmd i get ''fatal: could not create leading directories of 'kohya_ss.\setup.bat': Invalid argument'' can anyone help
do you know how feasible it is to install Stable Diffusion on a Mac? for example Mac Mini M2 Pro? Will it work by running it on a virtual machine to windows or will it be better via bootcamp? or is there a Mac direct version?
I like this guy, he's like a nerdy hugging bear man you wana hold and thank for making such cool helpful channel.
Best comment of the day 🥇
Sometimes i think he is AI
"Hugging face guy" lol
Down bad
Seems like a sweet guy
Dude I watch all your videos even if I already know what it's about just to bump your numbers. Same reason for this comment. You're killing it.
lets goooooo !!!!
Your guide on LORA and Checkpoint Model Training is a game-changer. The merging trick you shared is pure gold. Subscribed! 🙌
Of everything I learned, you blew my mind with the f2 trick
All your videos, and especially this, are invaluable to the AI community. Foremost the beginners, but also the enthusiasts that are kept up to date with the rapidly changing and improving technology. And explicitly worth mentioning: Love your enunciation! Your well-phrased, concise use of language and well prepared scripting makes your videos efficient and a pleasure to watch! Thank you so much for your efforts and contribution!
A really helpful and informative video. In addition, and I find this particularly important, your pronunciation is also really super. For me, as a non-native speaker, this is very important. In many videos, people just mumble to themselves, or you can not understand them so well for technical reasons. With your videos I found the first time someone I can really understand excellent and crystal clear. 😊
I may have missed the Lora training, I followed the Kohya installation and the checkpoint training. Thanks for all the excellent content, Olivio!
the taging changes a fair bit depending on if your trying to train a topic or a style or a specific person and it would be nice to show the differences there. i see in on another video they explained how you should tag everything you DONT want to be a part of your lora unless you are training for a single person or thing, and I wold like to see a more in depth dive into tagging and its effects
The tagging strategy is a pretty confusing. I've also seen tutorials that say you should tag only the things you DO want to be part of your lora/model. And some tutorials say to not use tags at all and let the AI figure it all out. And should you use a unique "keyword" tag to train it to that particular unique "keyword" to use in the prompts? I've tried different the approaches but didn't get any conclusive results. They were all pretty bad and didn't work as expected. 🤨
A video from Olivio covering just the tagging aspect with a demo for each approach would be awesome.
@@Beltalowda55 captions are what a model uses to try and predict what the picture should look like during training. Difference between predicted and actual image is used to adjust the model in the right direction. So it should be obvious that you want to describe everything that is not a part of the concept you're trying to learn, to help the model to make a correct prediction and narrow the direction for the learning. Keyword is added for the model to associate the concept with that keyword, that helps to preserve sanity.
It sould be noted though that the training process does not actually differentiate what you're trying to learn in any way. All it does is it tries to match a predicted image with the actual image. That means it can associate to your concept any other or multiple tags you used in unpredictable ways, basically splitting the concept between them. To mitigate that, no tag except a keyword should be used too frequently (in my experience 30-50% is the limit)
I don't even bother tagging anymore when training a person.
@@fritt_wastakenI wish you had some videos up to talk about this. I think I understand what you're saying but some of it is a little fuzzy. Thank you for the info nonetheless.
6:33 Lora and Models
17:50 Lora Training: files and folders
18:53 KoyaSS: the software to train the models
tnx
This is the most comprehensive model training tutorial I've found on TH-cam. Thank you for making this video!
Thanks Olivio, I was just thinking I needed a guide like this after getting tired of google colab. This will be invaluable - your're a super star!
Thank you very much. Your detailed introduction is very rare, and many of the details have been very helpful to me.
I am so grateful for your detailed tutorials, step by step instructions. A big thank you.
Te trouble with all these tutorials is that when I spend an entire day following them, the software I see does not ever seem to match the software in the video. When I reach an error screen, I never know why and I have no comprehension of what to do about it. It's quite daunting for an artist who thought the only thing he would have to master in life was using a pencil and a brush. :(
@@ScaryStoriesNYC I absolutly understand that. Software is developing so fast and you have to make tutorial more modular, to change the parts, whenever something changes in Software.
You take your time, talk with no rush but also go straight to the point. Thanks a lot, really helping me now that most tutorials are outdated
Agreed, he gets the pace just right. Not easy to do, some people take too much time with no benefit, some too fast and hard to follow. The balance here is just right 👌
24:08 the same when you said "512x512" is in the past, the same goes to folder structure, is not necessary to create 3 folders, only the image folder and the inside folder with your dataset, for model and log you can leave the main folder "Neon" and the files will be saved there with no problems, I discovered that by accident and no issues after.
You are a Rock Star... when it comes to Ai Images... Keep Up the amazing videos that you do man... I for one, do appreciate your vids!!!😊
One of your best videos so far!!! And currently my favorite TH-camr for SD.
Greetings to Vienna from Stuttgart, Germany.
Sascha
You are my hero today! ❤
I was searching for exact this topic weekly for last 2 months 😂 Last time was this morning. Not sure if I go to sleep now or give it a try 😅
Allez dormir heureux si vous etes fatigué. Vous serez frais le lendemain pour rêvé éveillé.
Thanks for the F2 rename files trick... I didn't know about it, you just saved me hours of renaming 😄
this video helped me understand a lot for another project im working on for langague model to use LoRAs to beef and nudge the actual model.
Excellent breakdown on understanding how to train Loras, thank you Olivio! I'm currently training with Fluxgym in Pinokio, makes it super easy.
Super cool! Please continue describe this topic for us in your next videos, I am waiting for those
Fantastic job Olivio! I really appreciate the work you do!
Wow wow wow what a man. I was waiting for this video
Hope you enjoy it and it help you a lot :)
@@OlivioSarikas thx oli
Lol I was just looking for an updated LoRA guide, great timing thank you!
Would love to see something like this but from the perspective of objects, not portraits.
Same
I've been waiting on a video like this for ages! Thank you so much
Really love your content, thank you for taking the time to produce and share it!
So cool, thanks for the indepth explanation 👍
Thank you for this awesome guide! Anything about regularization images would also be fantastic.
That tag manager is really cool, thanks for that.
Nice VDO as usual. You mentioned in the VDO that you will cover both Dreambooth & Dreamboot LoRA; however, there has been no Dreamboot LoRA covered. I am not sure whether both Dreambooth & Dreamboot LoRA will have the same process & pretrain model or not?
For art painting style training using Dreamboot LoRA, could you please suggest an appropriate Pretrained Model, and min. total step for each Epoch. I tried 100 Steps for sample 20 images (2000 in total) but when I tested with text prompts used in the training, the results were still far from original images. Also tried with LyCORIS/LoCon but with no success.
I would be very appreciated if you could demonstrate how to train a painting style like Van Gogh painting style (though it is already available in some models) showing that the result from prompts used in the training process can reproduce images used in the training.
Dreambooth doesn't make LORAs. It trains a model, then you can extract a LORA from that by comparing the original base model and the trained model. You can see the tabs for that in the video when he's looking over the Dreambooth GUI.
Danke! :)
Thanks Olivio for your many great explanation videos and tutorials and that you made the effort to really go through every step of the setup and first usage of kohya.
I needed exactly this to get to a starting point of the mysterious journey of training my own models/loras and now I can finally experiment on my own.
A little tip, when installing python make sure to check the box that adds pip to environment variables, for some reason mine did not by default.
mine also. I just uninstalled Python and reinstalled it over the Microsoft store. It will install pip correctly
how would I know if I installed it without checking the box?
@@plattepus Try installing and running SD, and watch for the error in the console, It will give you a pretty noticeable verbose error saying Pip is not installed. If so just run the python install again and make sure to tick the box to install pip in the menu. Hope that helps.
@@addermoth appreciate it man. Also, I'm only seeing "gui.bat" in the "kohya_ss" folder, and not "user gui.bat", but it still seems to launch it fine. And I can't tell if it installed cudnn, I put the cuddn_windows into kohya_ss, ran the command to install it, but it never said "done" or anything.
Thanks for this informative video. I have been trying to figure this out for awhile, so this was right on time
Just once I would like one of these tutorials to actually work and properly correspond to what I see when following the steps.
Thank you very very much! This lecture has answered a lot of my questions.
Thank you for guiding me with this excellent video!
nice ive been waiting for this video !
Took me forever to record this ;)
Good evening Olivio,
So I just followed your guide to create my first LORA but I find myself with 22 hours of training!
I must have made a mistake somewhere ;-) However, I have an RTX 4070 TI and a Core I9 13900k.
120 photos of the same model in high quality (FHD) half are faces and the other is the full length model.
I chose 10 steps, I think my concern comes from there. I'm still going to let it finish just out of curiosity.
Thank you for the guide which will have nevertheless been very useful to me.
Thank you so much sir, I love your YT channel ! =)
Thank you :)
Thanks a lot I get more information about training now 😊
This video was really needed. Thanks for that
mostly installation instructions very clear but not what I was expecting
The Kohya page and download isn't the same as in this video and the setup.bat file has been updated. It doesn't ask me the same questions as it does him in the video. How did you get it to work?
I would suggest to always select PNG over JPG since PNG will not lose image information, and the file is small enough anyway.
Correct, but don't use convert RGBA To RGB and transparent image to white Background, because it's make ur trigger background situation or etc, at your prompt be blank or just white or just template take a sample data.
Watch so many of your vids and I thought I've already subscribed to you. Until today I find out😅😅 Sub'ed anyways
Thank you for explaining very clearly all the concepts that apply to creating LORAs. I really appreciate the info in the video.
lovely vid, keep on these good works!
This is awesome, thanks. I think you did a great job on disseminating the critical information concisely, I wish more tutorials were done like this.
I have a question, how would training for an illustration style, using real photos for face reference differ, if at all?
Excellent Tutorial!!!!
18:41....👀I didn't know that trick 👍🏻 Thanks
THX! I learned so much! You are amazing
Thanks for the video Olivio! I got a runtime error when trying to train a lora and this is how i (think) i fixed it. When installing kohya, Olivio said that the questions prompts are correctly pre selected. This wasnt the case for me. The question prompt that has a selection for fp16 had to be selected during installation. It seemed that otherwise kohya was trying to use my CPU instead of my GPU. To fix my error, i deleted kohya install folder, reinstalled kohya, uninstalled and reinstalled torch when prompted, and selected fp16 when prompted. i started training a lora before writing this post and its still working so if you have a runtime error like i did, try this maybe. hope this helps!
Thank you so much! I wish this was specified in the video.
Did you have to restart the whole process?
Video is super simple compared to others :)
Thanks so much for this video! really appreciate it
By far the most thorough and easiest to understand guide. I followed someone else's this past week and they skipped.
The Cudda setup and explanation
The Folder Structuring setup explanation
The explanation of the different tabs, up until now I thought the Dreambooth was just a different way to train Lora's....anyway their guide was really confusing and missed a lot of stuff you cover.
was waiting on this update !
If you are like me and have a hard time finding high res photos of the target person you are trying to train, simply run all the images through face restoration. It will add all the detail you need for training. Hope this helps someone.
brilliant idea!
Can you explain a little further? You put a low res into img2img, set denoise to very low and tick restore face? Hires fix? I'm confused.
@@gohan2091 Topaz Gigapixel, or more like all in one tool: Topaz Photo AI. It has upscaler, sharpener, denoiser and face restoration all together. (not getting payed for this, i simply love to use this by myself)
i would suggest to use ligthroom to create keywords and export with filenames containing comma-separated keywords, and create captioning .txt files from that filenames with a simple python script 👍🏻 Much better keywording and organization. (smart albums, etc) Thanks
you are awesome. internet is getting better bcs of u 🎉
Fantastic tutorial Olivio! Love all your AI videos. Can you use LORA to train other objects and keep them consistent? Using your technique, I've been trying to create a LORA of a specific baseball glove and a specific helmet, and apply it to an image of a baseball player in stadium. Most of the outputs I get are quite monstrous like 10 finger alien gloves. I've given it a ton of data like close up/medium shots of the glove from multiple angles, medium and wide shot of someone wearing the glove. Still nothing works. Any thoughts on how to train a LORA to keep something like a glove consistent? I'm using using SD v1.5 pruned, photon v1, realistic vision v40 as the checkpoints in A1111.
Hello friend, is there going to be anything new on Lora training program for people? I haven't heard anything in a long time, maybe there are some new software chips
Hey Olivio! I love this tutorial! Do you have any plans to update it? Have a great day!
Great job Daddy :) I would like to thank you as someone who can explain such a difficult to understand subject so well.
Great video Olivio! As I am watching this, I wonder what goes into the "log" folder and "model" folder? I didn't see what those are for. Also, do we need regularization images? If so, how many?
amazing info bro tysm
24:15 When Olli said "The images are not in the images folder" I got really confused. Eventually I figured out that you still have to copy your source images into the new folder, otherwise it won't have anything to tag.
Me too
Great video! But I prefer to train Textual Inversion for characters (a specific person) in Automatic1111: it's way faster, simpler, I get the same result than a Lora and spend less HDD space and only 2 tokens for the prompt!
Is the likeness almost 95% to the original person training on TI? Is it better than lora for specific character only?
Do you have resources to learn more about this? I'm trying to train an anime styled OC
Do you have a video I can follow?
I would like to know more about that too if you would be kind enough to develop!
Is the step of "merge" the models also necessary with Lora? I got confused because I thought you were going to create a Lora but you ended up creating a Checkpoint.
I thought about that too...
yes, great vid, following everything easily. I went with LORA too (dreambooth Lora tab) and now have the model but not sure what to do next. Thanks!
Thank you! Can you create the video with the colab as well? don't know if the old ones around still work, this things change so fast!
Hey Olivio, great vidio. Really go deep and surface all in one. I have a question. The style model I want to use is Dreamlike Photorealistic, from the same people as realisticvision. In your case you trained with sd15 in Kohya-Dreambooth (with captions for training dataset), but your video is entitled Lora. With Lora I usually use class dataset and Training dataset. Not sure if that was intentional. But more to the point, should i train with SD15 and merge to Dreamlike, go apply that style? Should I also use captions? I am doing Lora,thought. Thanks
If I understand correctly, the merge is just a weighted average of the model weights, so the shapes of the models must be the same. So you should merge your LoRA with another LoRA that fits the style you like
My hero ! many thanks
YOU MAKE THE BEST AI VIDEOS!!!!!!!!!
Thank you so much for this ❤
Would definitely like a online workflow video
Great explanation!
Hi, thanks for your amazing videos. But I had a lot of problems. first, the setup: you didnt show us all the ssettings in the cmd, i was really confused, second the picture tagging: it just spit out some errors. and I am really stuck at this point. So I couldnt train any modells :(
Thank you for explaining
Is there a more up to date version of this that isn't a year old. Some of the tabs on this isn't the same as the current version and can be confusing.
Very nice on , now on my testcomputer it seems to work now..
What i am still missing is how to create a LoRA with this tool..
Perfect tutorial. Thanks for your work 😊🙏 I have noticed there is now also LoHa, have you tested, compared? Otherwise I’m very happy to use this workflow
The problem with Kohya is that is updating often and the guy like to move around and change things in installation, interface, parameters how to name things.
This is the 5th tutorial I try to follow... all of them have something different from the Kohya I have installed yesterday.
But all of them were usefull to puzzle out how to make it... storta work. I was able to make the first lora with bad results, but probably as I'm not able to load the reference pictures for some reason. It gives an error and do not use them.
when i press train. it shows this error " Not all images folders have proper name patterns "
Thank for sharing. 😁😁
Have you ever tried creating a MoE?
If so, how'd it go?
One thing I'd like to learn at some point is how to think of/define an "expert" in a project.
Great stuff, thanks
Great, but you skipped some confusing stuff....do I choose no distributed choice with a 4090 card, do I choose to optimize the script with dynomo, and which dynamo if so? Do I use DEEPSPEED, etc...I have no idea about these cryptic choices and what would be optimal... 😞 help?
...also, you have a 4090 so when you get to choose pick bf16. The f16 is for older cards, like my 1070
Hey Legend! I fell in love with your tutorials! I have a question: this is not working on my laptop and on my server as well :( I have 2080 RTX in both. Can this be the reason? In my server there are 2x2070... do you have any other method to train models for stable diffusion?
Hi Olivio, thanks for this guide, it's a bundle of usefull informations. I have a question: at 30:56, when there's your Kohya full screen, I read "Learning rate: 0.00001", the default should be 0.0001, is this correct?
Thanks. This video helped a lot. However i do have aproblem with Kohya. All the buttons are not working so cannot really do anything. The is a warnig about LD Library in Terminbal. Am i missing some files or what is going on?
I don't understand why an app can give better results with 10 images that are very similair but SD makes me look like a totally different person. even the tiktok filter gives better results with only 1 image (granted, it changes your look a little too much) but still gives a result where you at least look human. SD is much harder to train than I expected.
Thank you for the video! You mentioned you could do one with an online server, please, pleeeeeeease would you be so kind as to show us how to train a LORA step by step with RunPod ? 😭😭😭
Awesome video! Big thanks. I am starting to think that my hardware might be not up to this... whenever I run the kohya gui bat it just closes itself.
Thank you Olivio.
Thank you so much, Olivio! Is the training feature on Automatic1111 still broken?
Awesome video, please make video for Google colab also
yes plese make a video training checkpoint models using online servers
If I'm trying to train on an anime character should I use background removal to isolate her? I have many illustrations of the character with design elements around her rather than full color backgrounds and I'd like to use her LORA to do full color art and more. Thanks for the great guide!
Generated images will have solid white background, if you provide .png with transparent background.
When last time tried stickers it worked great.