Glad you like them! Thank you for the kind words :)

Very interesting, and a good tip!

Thank you! Have been having that exact error "Attempting to unscale FP16 gradients" too! Nice work and thanks again for sharing

@@McMartinLC the credit actually goes to Aitrepreneur who discovered this fix 🙂

You're very welcome!

Glad you enjoyed it! And thank you for taking the time to leave such a kind comment 😊

Glad you like it! Got a whole bunch of tutorials on the channel if you want to see more :)

You being able to use "Visions of Chaos" means I want to have your PC setup.

I've been playing around with that, it seems very powerful from what I've seen!

aha :)

ohhh ok thx for that! And if I got spaces in it...do I have to de- and reinstall everything to rename the folders?

oh crap, that would probably be my issue too, I'm going to have to move some things around

@@ZeroIQ2 just take the spaces out.. everything still hung together for me

@@DeliciousBrain12 just take the spaces out.. everything still hung together for me

Right!? From complicated installs to this. This is a real gamechanger.

Awesome, thank you so much for the kinds words! 💎 I'm using the rtx 3080 and it's fairly fast, however I would have wanted a little more vram, preferrably over 12gb. But unless you care about locally training Dreambooth models, something similar to 3080 is great. I get 10+ it/s. More with xformers, which is decently fast. The new 4070 and 4080 would probably be twice as fast. Stick to Nvidia though. Easier with Stable diffusion.

@@sebastiankamph Great, thanks for the reco! 4080 with 20 it/s would be .... fast, jesus!

Haha, glad you liked it 😊

run on cpu is how many time more than gpu ?

i'm on ubuntu and this worked for me. thank you!

my stable-diffusion-webui folder is empty 😮

You're most welcome! I'll make sure you always have the latest gossip.

Your purchase of 3090 is not a mistake! CPU is suuuuuuuuuuuuuper slow. 3090 godmode 🌟

@@sebastiankamph Still getting CUDA out of memory errors. When I try to use fp16, I also get an error "expected scalar type Half but found Float" - any ideas how to fix that? According to your sheet it should save a little more VRAM.

Homie how much you get that 3090 for and from where? Im about to buy a new desktop gpu

Nvidia stopped selling 3090 officially, you can still find what left on stock or buy used one

@@testales I saw in github about that scalar half thing getting fixed just hours ago. Try updating.

You're adding training to an existing model, yes. Then that new model is saved as a fresh model. The previous model is left as is.

No, the instance prompt shouldn't include the class prompt while training. That's set separate.

@@sebastiankamph how about at inference?

Thank you! Try this: kzfaq.info/get/bejne/sMd4e8-Uv8XRnmw.html

Yes! You're training the model to add your input images, which could be for example a person or an art style.

You're very welcome buddy!

That's an error I haven't seen before. Check the github for issues. Sorry man, it's a little buggy. Thanks for your support though ☺

You can try raising your weight in the prompt with (((firstnamelastname))). If it still doesn't give you anything, I'd try retraining with more steps. Usually, people have been requiring less training steps with this version, but your mileage may vary.

@@sebastiankamph thanks for the quick answer, will try that tomorrow!

Your Swedish is excellent! 😊

I run Windows native. My Linux days are far behind me. 5800x and a rtx 3080.

Got the same problem and now it has many new features in the new updated version. it seems that it needs VRAM to extract the model.

@@kimhan2615 welp, that sucks. Did you ever get it working? I don't know why, but i feel like it will probably never work for me lol

This is my preferred choice as well. It's the most reliable and in that sense, fastest.

resolved: use the recommended version of Python!

Glad you got it resolved!

Then the model has been overtrained. Try setting XY in the end of your prompt.

@@sebastiankamph Thank you, but I still dont quite understand. So the new CKPT-file should be used, right? When I do this, EVERY prompt is XY, even if XY is not part of the prompt text. For instance the prompt "cat" still ends up being a picture of XY. I redid the training with a completely different picture set and the problem is the same. There seems to be nothing in the CKPT-file but the new XY from the training images, so "cat", "car" or "landscape" always is XY.

Then you need to lower training steps

Did you solve the problem? The same situation. 11 incoming photos, 1100 steps, 0.000005 learning rate. Everything I enter is the same as the original photos.

@@morozov_creator I reduced the steps to ~20 per image. Now it doesn't overwhelm the rest of the prompt, but there are also way less similarities with the trained persons face

Crossing your 7 fingers on one hand SD-style?

its out now!

(I want to add my own dog pictures to the existing dogs in SD)

The ETA will be shown in the GUI and also in the shell.

I also want to know

It will sadly not start in gpu mode until more optimizations are made in development.

Automatic1111 is the best SD gui, you should be using that.

Why not make a HuggingFace Space?

THIS is what I want to know, too :D

Same issue on my end

@@SuperFunHappyTobi I solved it by manually deleting the dreambooth folder in the extensions folder. It worked afterwards. Well not exactly. I had to set the preview image generation to some high value because there's a bug and someone had to do the same with the checkpoint. But after this it worked and I successfully trained some models after that.

Great to hear it worked for you in the end!

In/outpainting is a little better in 1.5, but honestly, apart from that I'm not super impressed in 1.5 compared to 1.4.

Shout if it works, im on same (ti)...:P

@@stefanb7015 yup no go, tried everything in the ui. restarted a1111 after every attempt. Even cpu fails. I love the idea, so watching bugs on git and testing when I have time.

@@madebyrasa I'll get back to you in here when I've tried..

@@madebyrasa I get the same, tried everything. Also failed with the cpu (I'm on i9 and 32 RAM). It seems to still be a mess, but it will be sorted out in a couple of weeks. :P

@@madebyrasa works for me on 2080 ti, are you getting errors in your console? I'm on Linux so maybe it's a windows thing?

Haha, glad that got a laugh somewhere 😉

I don't think 8 will cut it. I still use Google Colab most of the time, takes about 30 minutes for that amount of images.

I don't think that's possible yet. But you can create 20 models with 20 prompt words and then merge them together to 1 model. I can't guarantee the results will be great with so many, but I've gotten it to work with fewer.

Need in models\dreambooth\YOUR_NET_NAME\db_config.json - change mixed_precision": "fp16" to "Accelerate()"

@@FlameWinds didn't work for me i keept getting denom = (exp_avg_sq.sqrt() / bias_correction2_sqrt).add_(eps) RuntimeError: CUDA out of memory. Tried to allocate 114.00 MiB (GPU 0; 16.00 GiB total capacity; 15.00 GiB already allocated; 0 bytes free; 15.18 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

nope.

@@ac3d657 lol

@@ZeroIQ2 I have 3060 too, let us know how it worked for you thanks

If 12 gb is not enough, how muck vram it needs? 20?

@@samanmonazami9156 I've having issues with it, it doesn't seem to be a memory issue though, I think I've got an issue with my setup.

same problem here

@@yatobatoshu8299 ill update if i find a fix, so far its random. im running on a 3090 too so it isnt gpu

@@yatobatoshu8299 multiple uninstalls and reinstalls. dont change a single setting and run it, sometimes works sometimes doesnt but im thinking will need to wait for update on it.

Someone mentioned setting the preview values higher than the training steps. Apparently some kind of bug. Might not work for you though but did for many others.

Honestly, since you're on a highspec computer, I'd just go to TheLastBen's colab, run a training there for 20 minutes and then import the .ckpt. Cpu is super slow.

Just delete your Dreambooth folder. Then git clone the github version you want.

@@sebastiankamph удалял, git pull восстанавливает то же что и было( скиньте кто-нибудь у кого в папке Extension еще лежит старая версия на 2 вкладки, а не на 8 =(

How would you inform it of your embedding? Didn't see such an option in this video. Does it mess up if you use your embedding while training a dreambooth model(overfitting)?

1. You can't if you forget 2. Merge

I'd recommend sticking with colab over cpu only. Cpu only is your absolute last resort.

I'd recommend google colab instead. Check TheLastBen

Me too.

Same here :\

Same :(

I figured it out - make sure the entire folder path that has stable-diffusion-webui has NO spaces in it.

@@anonymousmuskox1893 wow it works now! Thank you so much! 😊

Join us in Discord and ask there and we'll sort it out.

Don't join Discord. It's a ruse. They will kidnap you and use you as AI training material.

@@brexitgreens How did you know!?

@@sebastiankamph 😏😁♥️

Very nice! Happy you got it working and thank you for sharing what helped ⭐

Thanks heaps for this was trying to figure that one out

add it to what file?

@@anonymousmuskox1893 webui-user.bat

@@sebastiankamph Any Idea what version of python you have - this xformers is really causing me issues!

Hey! Hmmm, haven't seen that error before. Make sure you have Python 3.10.x installed and if you do, try reinstalling again. Heard people having issues but most have solved by reinstalling SD/Dreambooth.

I figured it out - make sure the entire folder path that has stable-diffusion-webui has NO spaces in it.

@@anonymousmuskox1893 you are godsent! Can you change the name without affecting the file structure? I was considering to remove the spaces but I thought it would break my installation

@@yatobatoshu8299 i just changed it and had no issue whatsoever

@@anonymousmuskox1893 THX 🙂. You are right :) I had space in my folder path. I deleted the space and now the installation works fine. 👍👍👍

same to me

@@tivadarczondi2514 fixed it. Reinstalled A1111 and followed this, voila

buuut... then when trying to run, it fails. get this: Error completing request Arguments: ('withMe', 'C:\DELEMAPPE\Trenemappe\Darkr\512x512\jpg', '', 'darkr', '', '', '', 1.0, 7.5, 40.0, 0, 512, False, True, 1, 1, 1, 1200, 1, True, 5e-06, False, 'constant', 0, False, 0.9, 0.999, 0.01, 1e-08, 1, 500, 0, 'fp16', True, '', False, True, True) {} Traceback (most recent call last): File "C:\Users\Darkr\stable-diffusion-webui\modules\ui.py", line 185, in f res = list(func(args, **kwargs)) File "C:\Users\Darkr\stable-diffusion-webui\webui.py", line 54, in f res = func(args, **kwargs) File "C:\Users\Darkr\stable-diffusion-webui\extensions\sd_dreambooth_extension\dreambooth\dreambooth.py", line 265, in start_training trained_steps = main(config) File "C:\Users\Darkr\stable-diffusion-webui\extensions\sd_dreambooth_extension\dreambooth\train_dreambooth.py", line 872, in main save_model = not global_step % args.save_embedding_every ZeroDivisionError: integer division or modulo by zero

@@birknygaard2944 Sorry, what is A1111? And how to reinstall it?

kzfaq.info/get/bejne/eq6Rf5hntZzKZYE.html

I know it wasn't possible a while back, but not sure if they've managed to sort that yet. I'd guess no, but it moves so fast now so who knows 😊

@@sebastiankamph Hey Sabastian :) thanks for responding, yeah,, im not sure as well,, If I learn of anything I will be sure to share that with you :) thanks again for all the great information,

Check my discord in channel description.

@@sebastiankamph Thanks, will do!

Great info, Johnathan! Thanks for sharing 🌟

I'll come back about that at some point after I've tried with my Quad-Core i5 3.7GHz. Might be on the discord though, because I doubt I'll remember to get back here. I'm ok with it taking time, if it works. Rather slow and working than failing super fast.

how do i do that?

​@@6EXAI you open a terminal window, then you cd ~/stable-diffusion-webui and in there you edit the requirements.txt file with for example nano requirements.txt. Then at the bottom you add a new line and type accelerate Save it, close it and restart stable diffusion.

THANK YOU

@@N0XQS Um this did not work for me, "accelerate" was already there near the top. I deleted that and placed it at the bottom, nothing, kept it in both locations, nothing. Left it in the same spot, still no tab popping up.

No, you will have to run on cpu only. Will take something like 24 hours. I would recommend running TheLastBen's colab in video description. Takes 30 minutes to get a .ckpt.

Don't I need a .CKPT file to add to the modules folder? I only see .BIN and .JSON files in the output folder

Also I'm getting only 100/1200 steps, then it says "training completed" which is weird, not sure if it's actually an error. I have an RTX 3090 so there shouldn't be any hardware problems

I've seen a bug where training can complete or bug out if the preview setting is 0 or lower than the training steps. Could be that.

Install Stable diffusion automatic1111 (Ultimate guide). If you want Dreambooth, use TheLastBen's colab for ease of use. Shivam's for a little better result.

@@sebastiankamph when I did it it was talking about Vae

More like there is now at least an option, if you want run something without a capable GPU.

Yeah, what devnull said. The option exists, but having a good gpu is the preferred choice.

@@devnull_ Do you know if I can run SD in a gtx 1050ti? It's a 2gb vram gpu but i couldn't find any information related with text to image generators

there's a 'cpu only' checkbox in the trraining parameters section

Sorry, no :( Go check TheLastBen's or Shivam's colab (see links in description). You can run CPU though, and that will take like 2 days. But colab is faster and better.

@@sebastiankamph ok man, thanks.

Different for different pics. I'd be happy to help you if you come to our Discord (see channel description) and ask there more specifically what image you want.