DREAMBOOTH: Easiest Way to Train an AI Model for Stable Diffusion
HTML-код
- Опубликовано: 12 фев 2023
- Update:
If you want to use the non standard model for 1.5 training then you can grab the name form hugging space such as XpucT/Deliberate, and use the word 'main" for the branch.
Instead of putting the trigger word in for the class you can use instead like "photo of a man" or "photo of a person" I've had better results with that.
I still recommend 1600 steps
This is the easiest way to train an AI model with or without a GUI for Stable Diffusion.
I wanted to try a similar thumbnail to Aitrepreneur as I really love the way they do thumbnails.
stable-diffusion-art.com/drea...
www.birme.net/
colab.research.google.com/git...
Other Great Resources:
/ @aitrepreneur
/ @nerdyrodent
/ @enigmatic_e
/ @tutorialsai
Awesome tutorial Russell!!!
This the only process what gave me results thx so much ur da goat!
thanks for this.
really a great video
no luck trying with a different base model though
Hey! Great tutorial. I wanted to ask in-depth about what I need to do with AI training and see if you can give me a hand. I've been generating 3D models of some characters and also making scenes with them. For example, one running. I've been looking for a way to create these scenes without having to 3D render each one. So, I've tried putting images of these characters in AI to make scenarios using them as a base, but I haven't been successful. What would you say is the best approach to solve this problem? Is it even possible to achieve what I'm asking with AI? Thanks a lot for your response.
We currently don't have anyways to render 3D objects with AI. Best option now it Blender or some other software and then AI on top really.
Hey! What a great video, Russel! Thank you!
Have a question: why Collab is better than just using Stable Diffusion on local files? Maybe I just didn't understand something in codes and so on, but it's look like similiar interfaces...
Using this collab is just easy for everyone regardless if you are running locally or not. Personally even though I run a 4090 in the few attempts I've tried training locally they have turned out worse than the collab.
@@RussellKlimas Got it! Thank you!
Thx for the vid Russell. At 05:50 I understand how to use a trigger word in prompting (I'm using Auto1111 locally), but when training my LoRA's, I don't understand where to _set_ the trigger word. I'm confused by what you're saying here that you went back and "I used the trigger word rkkgr". Where did you do that? Where \ how did you set it? Is the trigger word the Instance Prompt? I can see how you later -used_ that trigger, but not where you actually set it.
So originally in the collab it will say "photo of zwx toy" so replace "zwx toy" with your trigger word.
@@RussellKlimas Thanks ... so in the collab version you're referring to the "Instance prompt"? (@ 03:54) So Instance prompt = trigger word
@@1salacious only the part you replace. You need to specify if it's a drawing or photo of whatever and then what you want the trigger word to be
Does this technique only work for creating a person? Can i use this to create something like an Achitech design? Or maybe something like a normal map for skin texture?
You can do it for a style as well.
Thanks
Could it be that you are getting album covers because your class_prompt isn't saying that it is a person?
Bagaimana cara melatih dengan model yang berdeda ?
Misalnyaa aku ingin melatih dengan model Chilloutmix atau Deliberate ?
Apakah ada caranya . 😃
Terimakasih
5:55 where did you use the trigger word and what is the word exactly bc it is hard to understand. Thanks
I just replaced the generic triggerword "zwx toy" and "a toy" with my own trigger word. Then it processed it automatically after I hit play.
Any idea on why I have the "MessageError: RangeError: Maximum call stack size exceeded." error when uploading images for training?
edit: The issue was coming from Safari, can't upload images with safari... great.
Im running into the same issue.
what do you mean by you use this to train on to start a base? do you train it further on something else after this?
with the 1.5 model it's a great model to start with to try this out first. Then if you want to try and use other models to train with you can branch out from there. 1.5 is just very reliable.
@@RussellKlimas right ok thanks!
So when you train your own images does it go into their data set?
No.
great tutorial! my example images have been coming out looking nothing like the pics i used, i used 23 pictures, and i tried it at, 800, 1600 and 2300 and all have not produced any results that look like the pictures
I've run into a similar issue when trying to train lately with it as well. It's so annoying that the process changes so much. Going to try again right now.
@@RussellKlimas great 👍 I'd love to hear how it goes and if there are any work arounds
@@RussellKlimas any updates?
hi sir, I am from India and i was searching for this type of tutorials since long. thankgod i finnaly found your channel....do we have to charge for dreambooth?
to train models no you don't, just need a gmail account
You saved me more headache with your tutorial so tons of thanks to you for it, your tutorial's colab page was the only one that worked despite being old, the rest of them I tried only gave me errors and they never worked for me, it drove me crazy all day long trying to find a working way to teach the model how I look like and thankfully I stumbled upon your video which finally did it and managed to learn how I look like and the results are cool, so thanks, I subbed and liked the video.
On a side note, how can I get better results that would capture my likeness even better yet would remain flexible enough? (on the first try I used the default settings of 300 steps and let it learn on 20 photos of me, I have 145 photos in my original dataset tho), what are the best settings and base model for that?
For this kind of a flash pan process it's going to be difficult. You'll be better off using LORA's and training in OneTrainer or Kohya to get the most flexibility.
@@RussellKlimas I see, thanks, I might have to still consider finding a way to teach it via LORA/kohya then but its tough finding one that is working well, too many errors or bugs on everyone of them almost and some of the dependencies like Git and Python that are required for the process are containing malware or such so they are no option for me.
@@MrRandomPlays_1987 Do you mean some of the things you would need to download using Git and Python? If you obtain Git and Python through the official sources you definitely shouldn't have malware.
@@RussellKlimas Yeah, basically I scanned in 3 scanning sites their files from their official source and it detected in some of their scanners that it is malicious/contain malware
@@MrRandomPlays_1987 Hmm I don't know what scanning sites are but here are the official sites for those
git-scm.com/downloads
www.python.org/downloads/
Python wise I prefer 3.1110
There is something missing here imho. Where did the tags come from? Is SD adding these images in to it's premade models then? Sorry for wrong terminology here. I'm still trying to figure out the architecture behind SD.
What do you mean by tags?Yes you are training the images into a model that's already been made.
@@RussellKlimas He's talking about picture data description, segmentation, etc. Usually a txt2img model is trained based on text-image pairs.
@@RussellKlimas don’t take me wrong. Can we do this without a model from scratch?
@@nithins3648 You would need millions of images to make your own model and would need an insane graphics card to do so like an a6000
is it possible to use a model from civitai or some other external site? Hugging face doesn't have the best models.
I don't know 100% I would ask the maker of the repo here! They are pretty responsive. stable-diffusion-art.com/dreambooth/
I think you need to merge your model with other instead of.. Both are checkpoints of sd1.5, right?
How fix "404 Client Error: Not Found for url (name of the model git)"? only work fine stable diffussion model to me.
check the comments here stable-diffusion-art.com/dreambooth/ and ask him questions. He will have better answers than me.
People always use faces to demonstrate this process, but it'd work for anything right? Power Rangers, cactus plants, fish, buildings, etc?
Yes should work for anything
Anything you have enough photos of to train stuff
I like to do it with dress like pants ,shirts , sunglasses is it possible
If am doing dress which model should i use please reply ❤
Hi, I have a problem. When I click the play button, it says that I have a FetchError. What do I do?
Hmmm I'm not certain. Reach out to the creator at the stable diffusion art link and ask on their blog page. They are pretty quick to respond and have helped me out before.
Really thanks for your videos and time :) A lot of peolpe saying the latest version of Dreambooth is broken. Is that really real? :(
It's been a huge pain in the butt in the latest version. I just used collab fast dreambooth and with 25 images and 3000 steps was able to train a custom model no problem. Apparently also works at 1500 steps and 15 images but I have not verified.
@@RussellKlimas does fast dreambooth colab produce the same quality or is it worse too?
@@donutello_ fast dreambooth collab is producing good results right now
Oh, so now it costs $5 to get access to the Collab? fuck that shit. lmao
It does if you use his. If you check the link for the collab in the description you good to go.
i am getting a "OutOfMemoryError: CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 14.75
GiB total capacity; 8.17 GiB already allocated; 10.81 MiB free; 8.31 GiB
reserved in total by PyTorch) If reserved memory is >> allocated memory try
setting max_split_size_mb to avoid fragmentation. See documentation for Memory
Management and PYTORCH_CUDA_ALLOC_CONF" error,can you help me?
You used all the available free GPU you have on your account with google collab. You can try running with CPU or make a new gmail and use the credits there. I have created several for this reason.
Hi, Does anyone know how to fix this error?
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchdata 0.6.1 requires torch==2.0.1, but you have torch 2.0.0+cu118 which is incompatible.
torchtext 0.15.2 requires torch==2.0.1, but you have torch 2.0.0+cu118 which is incompatible.
Successfully installed torch-2.0.0+cu118 torchaudio-2.0.1+cu118 torchvision-0.15.1+cu118
WARNING: The following packages were previously imported in this runtime:
[nvfuser,torch]
You must restart the runtime in order to use newly installed versions.
If you haven't resolved it yet, have you tried uninstalling torch completely before reinstalling?
I am also having this issue. I have uninstalled and reinstalled torch about 18 times now from many different sources, EVERYTHING IS UP TO DATE. And somehow, it still. Doesn't. WORK.
Did you ever manage to get this fixed?
If so...
... H O W ???
Looks like youve got the wrong version of torch installed you need the version shown above, 2.0.1
Is it possible to train it based on videos? specifically video tutorials?
Not this one, I mean in general, dont really see a AI that train based of videos
Models technically work either way. They can work off the same models. It's comes down to the process on how the video is actually made for it to matter. Similar to Gen 2. Or at least to my understanding.
So I'm following this but was asked to pay $5 to get access to the collab dreambooth. Now Google wants another $13 because my GPU type is not available...am I getting scammed here or do I have to pay to get this working?
I found the solution. If you get the GPU not available error then go to runtime-change runtime type and select hardware type none. Now I'm stuck 'no training weights directory found' :(
@@philjones8815 what model are you using? Depending on the model I get that error too. I know that Realistic Vision and Revanimated work if instead of FP16 you put main.
@@RussellKlimas I was using the SD 1.5 model but I'll try using Realistic Vision without fp16. Thank you so much for the reply.
I had to use the original model but 'compile xformers' in order for this process to work, even when I had xformers installed. Great tutorial Russell, and I hope people find my hair pulling experience helpful in achieving their goals.
Didn't work for me :/ Kept getting "RangeError: Maximum call stack size exceeded."
Hmmmm I've never run into that error before. Wish I could be of more help.
Can we do this process locally?
Probably? If you connect it to your graphics card you could. Otherwise you can use the automatic 1111 UI with the dreambooth extension but I find that confusing.
you'll need a beefy GPU for it
How would I then use this on controllnet in Colab?
If the collab has controlnet you would use it in the same way you would use collab without controlnet. It makes no difference.
@@RussellKlimas Sorry, I don't follow. Let me rephrase my question to be a little more clear. I've used Dreambooth to train a model. I can add my promt into the the promt input of the Dreambooth interface on Colab and they come out looking great. But now I want to be able to create images with this model on the Controlnet interface on Colab, so I can also use OpenPose editor with it. How do I load the Dreambooth model onto Controlnet?
@@anaximander9215 If you're talking about making a controlnet model I don't know how to do that. Different process.
@@RussellKlimas No, I'm not talking about making a controlnet model, I'm talking about using the model I created in Dreambooth in controlnet. At 8:00 you say "to keep the model within that so you can call it when you want to generate". That's what I'm trying to figure out how to do. How do I call the model in controlnet to use it to generate there? I'm sure that's a very basic question, but I've never used controllnet until now, so don't know how.
45 Gigs????
stable diffusion art is now a pay site
but if you check the link in the description for the collab still works
It no more works Russell, tried hard but no success dear.@@RussellKlimas
Also can you record it again for our ease? I mean as an update to current settings.
I was excited but got a error talking about "no training weights" im seriously annoyed that I get errors on every way of trying this shit
It's most likely due to the model you are trying to call to train. I've run into that error before. Definitely hit up the guy on stable diffusion art website. I wanted to use RealisticVision and he set it up to make it work for me.
@@RussellKlimas broo its doing it on every single freaking one I choose. ive put 9 so far like BRUH
@@theaccount4100 Try asking the stable diffusion art guy what's wrong and sharing your errors with him. Did you try just the basic of what's in there first?
@@RussellKlimas no I found out most of the ones dont work. I found a couple that do but its not a bug on my part its just the shit barely works correct. I trained a model with 600+ pics and it looks like shit an dont even show the face right. I tried merging checkpoints to make the body better. Its def a no go. my thing is like if deepfake computer programs can use data sets easily why is it so complicated to do it on a single photo? They made it hard on purpose.
@@theaccount4100 You could be overtraining. I never use more then 30 and 1600 steps.
1,000 steps for each photo, but then you said 2,500? should I be doing 25,000 steps, or did you mean 100 steps per photo?
100 steps per photo, though I've had good success at 1600 steps at long as the amount of photos is over like 14
how to train on other models?
put in model name in the begining
@@GuitarWithMe100 True, but I think people is confusedly calling models to checkpoints.
I got all deformed, lol
Why google drive?
Shouldn’t this be on your hard drive?
I mean you can run it locally if you want but this way it just easier.
45gigas?!?
that's insane
I am truly sorry regarding my comment, but NOBODY i say NOONE is actually showing how to train your own model. Alway going to some database that there are tons of models, that you need to waste literally days to get what you need and want. HOW TO TRAIN YOUR OWN MODEL FROM SCRATCH ?! Like literally not using someone else's preferences for body, face, nose, eyes, hair, skin, legs.... etc. Is there ANY video that makes sense for people totally new in this, besides "click here, go there... you are done". No i am not. I didn't get a thing regarding what i am looking for. :(((
Training your own model from scratch will take millions of images, and super high end graphics card and at this point in time around $10,000. That is something that I do not have the financial capability nor hardware to do.
@@RussellKlimas I see. I have the resource but no idea where to start. Too much conflict and basic info. I was looking for deep ML in that field. Thank you for your time! Appreciated!
they've locked the colab behind a paywall
my version still loads up no problem. Hence why it's in the description.
can I train an object instead of real people?
yes. same process.
Not always, is not working on my wooden toy dog.
"click this, click that."
Bruh, explain what it all does, Jesus.
This is more of how to, not how it works.
perhaps he could also stroke your hair as he explains it all for you, maybe bring you tea?