ULTIMATE FREE TEXTUAL INVERSION In Stable Diffusion! Your FACE INSIDE ALL MODELS!

[@Innere_Stille]

for everyone getting the error: completet training after 0 steps. Set the batch size to 1 and increase the gradient accumulation instead. worked for me :)

[@ImAnOcean]

worked for me too! upvote this hero, a lot of people are asking

[@BMXprorider98]

i tried it and sadly it still gives out the same message :/. I spent hours preparing all the images.. now i'm at a standstill. Looked it up on google but not much has been found, only a couple github posts and a reddit group which is private, no actual solution. I have a RTX 3060. any help?

[@charltonho]

I’m stuck. Didn’t work for me

[@krzysztofczarnecki8238]

@@charltonho closing Stable Diffusion starting it again (the actual command line program, not the browser page) helped in my case. The same if you try to generate an image that is too big, it will return an error but not free the VRAM automatically. Then you try the default size again it still won't generate even though it could before, and you have to restart it.

[@user-xg1jv4rw7e]

​@@krzysztofczarnecki8238yes it worked for me in vram error but my stable diffusion lags my PC and sometimes even freezes it even when not generating any images, do you have any idea on how to fix this?

[@07m07]

Just want to say, the way you word your explanations and go over them with examples and hypotheticals when necessary is really good. It's rare to run into anyone, especially on KZbin, who presents relatively complex steps and concepts with such clarity.

[@jokinglemon007]

Well, afterall, he's an AI

[@SnoochyB]

I train multiple embeddings daily. My style differs from yours, but I do not doubt yours will result in an embedding with quality likeness. There are def. multiple ways to train an embedding and get a good result. The Mila embedding featured in this video was made by me. Thanks for featuring it. I'd be honored if you would link it in your video description.

[@Aitrepreneur]

I loved your Mila and emma stone embedding, very accurate and precise! I added the link to your embedding in the video ;)

[@SnoochyB]

@@Aitrepreneur Thank you very much. It has been an adventure of learning, and you were there with me at the very beginning. I really appreciate all the information you provide to the AI community! Please keep up the fantastic work!

[@digidope]

Please do share how you train your embedibgs too.

[@MichaelTuckerLA]

@@SnoochyB Yes, would love to hear your workflow as well

[@franzkekko2387]

@@Aitrepreneur Am I the only one who gets an error after preparing dataset? “Error completing request”

[@Azroy6229]

This guy is so very helpful with great and detail explanation make me easy to understand. TQ so much!

[@Wastedtrait]

one of the most useful stable diffusion videos ive seen to date. Genrating a graph of all the different models is a game changer

[@sahanhasret]

There is no preprocess tab in my interface, what should I do?

[@mauvaissigne]

i have the same problem. did you ever figure it out

[@bobcoburn5238]

​@@mauvaissigne seems they moved it to the extras tab under "caption".

[@mauvaissigne]

@@bobcoburn5238 I know use comfy UI. It’s much better

[@CHARIOTangler]

This is absolutely one of the most useful and well-presented stable diffusion tutorials in youtube history. Thank you for being great at what you do🙏🙏🙏.

[@CoqueTornado]

I am a fan too

[@LAyeredx]

information overload tho @@CoqueTornado

[@neamedia]

Great insight on "textual inversion" - but how about an updated Dreambooth tutorial? With the latest Auto1111 version, and how to optimally train a face considering all the settings - maybe also doing it on rundpod?

[@Neko-et2bb]

I have had zero luck with Dreambooth since the update to the UI. I am sure it is user error

[@neamedia]

@@Neko-et2bb Me too... I can't figure out a good settings/concept input constellation, so I keep under- or overtraining the model and the outcome looks horrific... Also, I can't tell if it is better to train a face with DB and then merge with the protegen ckpt - or actually use textual inversion method shown in this tutorial... well I guess I'll do some more trial and error testing loops

[@mikemenders]

For the time being we have to wait for this, because the Dreambooth extension is being reworked because of LORA. I think it will be after that.

[@pedroenmanuelfuenteszambra8873]

I finally understand embeddings!!! This is awesome!!!! AI is the future for god's sake. I could train an embedding just in cool watercolor backgrounds and make comics or something!

[@PolarBearon]

Hey, if any of you struggle to get this working, like i did. (getting random unrelated images as a result of the training). The solution for me was to update my xformers. Apperantly the version i was running had a bug related to the "Use cross attention optimizations while training" option he told us to activate in the video. So either turn that off(which will slow you down since you'll use twice the vram as compared to it being on), or just update your xformers. That fixed it for me. Before i fixed it the training just yielded random images of people, trees, traffic signs, all kinds of stuff. Certainly not my subject.

[@matheusmf5863]

daaaaam thanks a lot, I was very WTF about that

[@matheusmf5863]

I unchecked this option but didn't solve :C, now i'll try update xformes

[@TheJefffree]

how do you update xformers? i tried looking it up online but had trouble. Any help much appreciated. Thank you. This was a GREAT tip.

[@XaosGamevids]

Even after updating xformers to 0.0.21 i still get some random resulting images of trees and backgrounds, any other tips ?

[@rainniu]

Huge. Thank you for the video. There are a lot of text-based guides out there, but a picture is worth a thousand words. I appreciate you, brother.

[@RachelHardy]

ii followed this tutorial with 50 reference images, on a 3070, and it's taken an hour to reach 2% finished and 54 steps. The estimated time keeps going up. Am I doing something wrong, or is my computer just not powerful enough?

[@FullSailGameDev]

Great video always learn something new for my workflow. I'd love a video on how to successfully train a style. I have great success when training a subject but when I try styles it always becomes a mess. Any suggestions on current guides focused on style or if you could make a video on it? Everyone always does subject videos.

[@euchale]

replying so this moves further up.

[@MrCleansweep]

Same, would love to learn this.

[@brunofernandesleandro9977]

Same here!

[@basiccomponents]

I would love this too!

[@FrankWildOfficial]

Hi! What would happen if used images of 2 different subjects(for example 2 diffrent girls)? I mean, let's say 10 images of Jenna and 10 images of Mila. Will the result look real as a woman that will look like mix between these two? Also, If this is not possible with embedings, is it possible with dreambooth? Thank You!

[@stephenirving9846]

This would be a fun experiment to do. I'd imagine you'd end up with images that look strongly like Jenna or strongly like Mila when using that textual inversion. What I would do is I would train a model of Jenna, Train a model of Mila, then use both of these textual inversion to generate 10 images of a new person. Then train a new textual inversion of that combo that you like.

[@Aitrepreneur]

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

[@franzkekko2387]

Am I the only one who gets an error after preparing dataset? “Error completing request”

[@sharkeys]

Would you consider showing how a combination of a trained CKPT dreambooth model and a trained embedding help with quality and coherence. Oh and maybe a video on the new Ultimate upscaler that just came out?

[@nikgrid]

K I am getting this error what does it mean? I have a RTX 3080 10gb vram

[@sharkeys]

@@franzkekko2387 check firstly that all the files are jpg or PNG. Ideally convert them to be all the same of either format. Double check the dimensions, check if you have any subfolders within that dataset folder. If it still doesn't work move your data set to c:dataset or something similar might be string length or similar

[@franzkekko2387]

@@sharkeys it’s all set… I don’t know what to do…

[@Roughneck7712]

This was great! I’ve been wrestling with embeddings for a while and you answered MANY of my questions. Thank you!

[@joergroser2536]

So we went from Textual Inversion training to Hypernetworks, Dreambooth collabs to local Dreambooth, and then back to TI :) As you said, TI embeddings are far more flexible and lightweight than HN or DB checkpoint files, and you can even mix TI embeddings together in a single prompt.

[@heyheymaizo]

At 6:53, how long does it usually take for the preprocessing to finish? After I clicked the preprocess button, it shows 'waiting...' and it's been like that for almost an hour now 😅. I checked and no files were created in the 'processed' folder.

[@Blootube1]

I have the same problem. Did you find a solution? Thanks.

[@Blootube1]

Were you using google colab like me?

[@heyheymaizo]

​@@Blootube1I gave up and never tried again haha. Yes I was using colab. When it was still free to use.

[@KrypticSloth]

So I’m at the training portion of the video but I am getting a memory error every time I try and run “Train Embedding” even if I have the batch and gradient down to 1 and 1. Why would that be? Could it be because I’m using too many images? I’d rather not set this up on a run pod and just do it on my own PC. I have 8GB of VRAM but it says that I have 7.06GB Reserved for PyTorch.

[@Aleksandrsvideo]

I have same problem

[@VNDROID]

yeah, same here exactly, it´s trying to allocate 1GB of RAM but cant do it because theres 6.5 GB reserved for pytorch

[@Quwerrr]

Same problem, xformers and medvram cant help it, tried some code twiks nothing changed

[@bobjohnson9354]

same problem. I've seen at least 50 persons with this issue, no one knows how to fix. Reddit, Huggingface etc, no answer. Crazy

[@Quwerrr]

@@bobjohnson9354 I deleted "--no-half" from webui-user.bat and now everything working. maybe it can help you too.

[@jordisales8351]

The best video I've seen on this specific subject so far. Thank you!

[@CarstenHensch]

Danke!

[@twistedream13]

I have followed all the step but when I train the embedding it looks like it starts but it says "Training finished at 0 steps." and there is nothing in the images folder. Any suggestions? the only things I did differently are I have 27 images, and I used 9 batch size and 3 gradient.

[@ashghebranious]

I get the same issue, Only using 8 photos myself so doing a batch of 8 and grad of 1, but zero produced in the embedding images folder

[@Innere_Stille]

same here

[@MrEnyecz]

Just note: lower learning rate is not a solution against overfitting. Overfitting (you said overlearning) is because you had too few images and the model was bent too much to the examples it got. If you start lowering the learning rate, the overfitting just get further away, but it will come (actually, it may even make the best solution before overfitting worse than otherwise). Lower learning rate can help, if there is instability in the learning (spikes in the loss function), but against overfitting, the only real solution is a bigger training dataset, i.e. more images in our case. Yes, there are other tricks like using data augmentation, regularization and/or dropout layers, but this is not what you can control from A1111 (and it is already used behind the curtains). Otherwise, thanks for the video, it was very useful!

[@gordonbrinkmann]

When you want to see the influence of the different steps with an X/Y plot, you can save quite a lot of time generating the plot if you don't use a placeholder word like "NUMBER", because it will generate a complete column with "NUMBER" used in the prompt although you have no embedding called wedmz-NUMBER. Just put wedmz-500 in the prompt and use 500,1000,1500,2000,2500,3000 for the X replacement, this will give you exactly what you want.

[@mcworks6822]

Interesting

[@joachim595]

Exactly. Don’t know why he said that stuff with NUMBER.

[@Todestelzer]

Ah nice trick. So 500 is the S/R word no need to use a Spezial word. Have to try it out!

[@CoqueTornado]

true, there is no need of writing NUMBER for the X

[@Todestelzer]

Ok. The only thing no one ever mentioned is that S/R only works in the Y field. In X it does nothing 😹

[@KnutNukem]

When all the zeros after the decimal point make your head spin, you also can use for e.g. 0.01 1e-2, or for 0.00005 5e-5. Xe+/-n stands for x*10^+/-n

[@AurelManea]

much better, thank you....

[@Kickassdave]

There are a few tutorials. This is HANDS DOWN, the best. I spent the last 2 days trying to figure this thing out. Dreambooth extension seems to have a lot of bugs and essentially, so long as you know what you're doing here, this can achieve the same type of results. Thank you.

[@ericbalfour5691]

i have try on runpod because with my pc, it say 6h for render, but i have the error RuntimeError: CUDA error: invalid argument CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1.

[@magenta6]

Thanks for this AiTreprenur! Very timely for me. I've had to disable Dreambooth because it caused an error that munted txt2img previews. Appreciate the time it takes to produce, edit and upload these quality tutorials.

[@GolpokothokRaktim]

You're enticing me to delve into ai and I am searching AI since last month. Recently I used BW for making images and I am amazed by the results

[@MrPlasmo]

TLDR: 1) Props to Aitrepreneur for finally someone going thru all the steps to create an imbedding. 2) Dreambooth training is still MUCH better than Embeddings unfortunately.

[@Thozi1976]

and another question: what would happen if I train the embedding on a model that was itself trained on SD 1_5 (like protogen e.g.) - wouldn't that be the same result?

[@PedroSantos-ty8sn]

RuntimeError: CUDA out of memory. Tried to allocate 128.00 MiB (GPU 0; 6.00 GiB total capacity; 5.01 GiB already allocated; 0 bytes free; 5.22 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 my gpu is only 6GB there some way to make it work? 😿😿😿😿😿😿😿😿

[@dkf-nl1703]

Set batch size to 1 and set gradient accumulation steps to 1. If you still get the error, then adjust webui-user.bat and add --medvram (that's two dashes) or even --lowvram to the COMMANDLINE_ARGS and then restart the webui with that adapted script. Hope that helps!

[@Modozenmusic]

oh i have the same problem

[@valhalla0112]

@@dkf-nl1703 i tried but no result, still cuda out of memory. How about @Pedro Santos, did it work?

[@lefourbe5596]

was waiting for this one thx :) ... however sadly ... theses settings give me an error : RuntimeError: CUDA error: invalid argument CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect. [+] bitsandbytes version 0.35.0 installed. [+] diffusers version 0.10.2 installed. [+] transformers version 4.25.1 installed. [+] xformers version 0.0.14.dev0 installed. [+] torch version 1.12.1+cu113 installed. [+] torchvision version 0.13.1+cu113 installed. (torchvision changed from 1.12.1 to 0.13.1, K is running 0.12.1, this might be the issue ?) (it is still working on one of my older version of the webui so i can workaround)

[@jamessurine3173]

The one issue I see with this recipe is that the learning rate step curve is fixed to the steps (e.g. 0.05:10, 0.02:20, 0.01:60, 0.005:200, 0.002:500, 0.001:3000, 0.0005) but the step rate is not fixed as it is dependent on the batch size and Gradient accumulation steps which is suggested to equal the number of images. I did not end up with good results quickly when using higher batch sizes and gradient accumulations when using larger number of images. I think the learning rate curve prescribed seems to work best with a batch size of 1 and gradient accumulation size of 1. This learning rate step curve probably needs to be adjusted when using batch sizes and gradient accumulations other than one.

[@reizero246]

Unfortunately the learning rate depends of much factors, not only number of images, also the quality of them, even your hardware can affect. So you have 2 possible scenarios, low learning rate and long times, or stay vigilant to the training and stop and remake the learning rate as its mentioned at the end of the video. If you use batch size and gradient accumulation size of 1, and you have for example 10 images, you need 10 steps to make a epoch, so your training is slower than making 10x1 or 5x2, or even 2x5

[@danieledallavalle5116]

I don't know if you are aware of how much skilled you are to teach things to people, thank you!

[@JoeHippeaw]

18:40 to continue the training further, Do I need to use only the same photo set as the previous training? or Can I add some new photos, or Is it possible to use entirely new photos on top of the trained embedding?

[@Antares2]

How does this compare to Lora? I've been creating Lora files with the Kohya-ss GUI to get new faces into SD, with very good results. I use 10-30 images and can make a decent Lora in about 10-15 minutes. The most time consuming work is actually finding the high quality source images. The training is very quick. The Lora files are a bit bigger, mine (I've made five so far) are about 9mb, which is still quite tiny. (though I've seen some people make 100mb+ Lora, even though I have no idea how they get the files so big) Textual inversion seemed a bit confusing to me, and in my case redundant as I get faster results with Lora, if it actually takes hours to do the training with textual inversion like a lot of people are saying. I haven't tried it yet, so I don't know. Also, with Lora, I can combine two faces fairly successfully, creating a mix between two people. Can textual inversion also do this?

[@ivandeus]

This is much much better than original documentations!!!! Thank you!

[@doingtime20]

Guys if you get an error that looks like this: "No operator found for `memory_efficient_attention_backward` with inputs" you may need to disable "--xformers" from the bat file.

[@Maulana-Al-Bakrichod]

NO matter whichever tutorial I follow. My Embeddings are nothing like what I want. Surprisingly, if I create a new CPKT or SAFETENSOR file, it gives me to desired result even when merged with other models. I AM GOING CRAZY NUTS OVER THIS! HELLLLLLLPPPPPPP MEEEEE SOMEONE PLEASE!

[@DatDood369]

I'm running a RTX 4080 OC, following everything at 17:29, except I changed my max steps to 2500, and it still says it'll take about 12 hours. Any help?

[@ilyakostin5676]

Any solution? RTX 3070TI 57-60 hours

[@happycollapse6348]

30 hours on RTX 4070, I don't understand :(

[@bolatmukashev2830]

Create embedding: "Number of vectors per tokens" this is how many tokens will be used during training. Tokens are taken from the Initialization text. 1 word - 1 token, comma - 1 token, space - 0 tokens.

[@zimnelredoran9985]

Hi, thanks a lot for this video! It's awesome the versatility that embeds have. I followed the vid step by step and I'm having the following error while trying to train the embed: "AssertionError: No inf checks were recorded for this optimizer". It stops at step 1, 2 and step 6 sometimes, but it always gets the same error. I tried lowering batches to 1 (using 15 images, nvidia rtx3060 8gig, Automatic1111) and lowering steps to 1500 and still the error persists. If I find out, I'll update.

[@gohan661]

Same!

[@gohan661]

ok I got it, when creating the text file he deletes the word 'a' but doesn't say it so it's easily missed. If you make sure it's 'a photo of [name], [filewords]' then it should work. MAKE SURE YOU DELETE THE PAGEBREAK/RETURN at the end of the line or the same error appears

[@zimnelredoran9985]

@@gohan661 Thank you!!

[@CtznofEternity]

@@gohan661 This comment needs to be pinned.

[@matheusmf5863]

@@gohan661 you sir deserve more thumbs up

[@cybermad64]

Woah thanks ! Sooo much info!!! I've been trying to train my own face profil for a week without proper result! Definitly gonna try it again tonight :)

[@cybermad64]

the X,Y plot is definitly a revolution on my side

[@TINTO_BRO]

How was it? Did you get good results? Because I didn't.

[@cybermad64]

@@TINTO_BRO I haven't yet :(

[@lefourbe5596]

@@cybermad64 i did train 3 3D character on the last version. 20 to 30 images each. With prior preservation disabled. 1 is fine, 1 is just okay, the last complicated one sucks.

[@delarosa6254]

your timing is perfect , i was about to ask if there is a video for this

[@Rewe4life]

I cant find the Preprocess section on my stable diffusion webui. do you know where it has been moved to?

[@mauvaissigne]

i have the same problem. did you ever figure it out

[@Rewe4life]

@@mauvaissigneyes, you can find it under extras -> caption

[@mauvaissigne]

@@Rewe4life hello, I found how to process the images under Extras-batch from directory-then I put the location of the image for the input-made a new folder for the output. But for the captions, I selected BLIP but the caption text file is not in my output/processed folder. Any suggestions?? Thanks

[@Rewe4life]

@@mauvaissigne it worked for me. Do you have discord so that we are able to send us screenshots to help each other? I'd look later into that deeper as I have to start my "GPU Server" in order to see the UI.

[@Beanyvids]

Question - instead of leaving the description as "a woman", shouldn't you change it to specify who it is, ie "Wednesday", or perhaps even the name of your textual inversion itself ("wedmz wearing a school uniform" etc)? To indicate it is the subject of the textual inversion? This part confused me (at about 7:50)

[@Beanyvids]

nevermind, I see now how it works when I got to the actual training step for my model

[@mrskull3607]

hello, thanx like always 4 your videos, i have a question when i start the train embadding i got this "Training finished at 0 steps. and nothing happen 😭

[@sebastianjung3168]

Try unchecking "Move VAE and CLIP to RAM when training if possible. Saves VRAM." in the settings option.

[@mrskull3607]

@@sebastianjung3168 oh thanx man, its WOOOOOOORK ITS WOOORK

[@antonioterra5436]

@@sebastianjung3168 Didn't work for me, still the same error.

[@marcelodesaellwanger4287]

@@antonioterra5436 same

[@ramongranados]

Hello, informative as ever. Should this work with a 3060ti? I keep running into the CUDA error about not having enough reserved memory (20 images @ batch 5 by gradient 4). Thank you for your time.

[@lefourbe5596]

Try batch size 2 or even 1if so. The higher the number, the more vram you'll need.

[@terraincognitagaming]

@@lefourbe5596 I have 12GB (4070ti) and even I cant do any combination. I get the same cuda error every single time.

[@JackiJ0nes]

i set up the settings exactly as you told to but the estimated time is round about 700 hours, what could be wrong?

[@1DusDB]

4:16 Should we keep that recomendation for train characters of including also photos showing half-body and others full-body?

[@ilyakostin5676]

am i doing something wrong? i have rtx 3070 ti and according to the video the training will take an hour, but for me it takes like... 36-40 hours

[@happycollapse6348]

Same for me , 28 hours on RTX 4070, with the same number of images. Did you find the problem ?

[@kaant21]

​@@happycollapse6348becuase you are out of vram it use normal ram. I was using 3070ti now i get a 3090 it took around 1 hour while on 3070ti its 30+ hour

[@joseffritzl8379]

This is a fantastic tutorial and I think it would work wonderfully for me...except my GPU is only 4GB VRAM and I keep running out of memory...either that or some other python CUDA setting is out of whack and frankly I'm not a programmer to understand which. Thank you for the video though it's straightforward and the whole process worked smooth right up until it didn't on my end!

[@ashtkon]

At 19:10 when I select a different embedding that was created and try to re run the training it just says completed without doing anything?

[@cyberfalcon2872]

Remember that max_steps is multiplied by batch_size and gradient_accumulation So 3000 max_steps with 11 batch_size equals 33000 max_steps with 1 batch_size

[@sharkeys]

so would it be faster to use these settings on a 3090 (currently using batch 8, gradient 1) Batch size 4, Gradient accumulation steps 2

[@cyberfalcon2872]

@@sharkeys I belive batch_size gives you slightly better results than gradient, but gradient saves more memory. Do the most batch_size your GPU can handle and then increase gradient to match your image pool (when multiplied by batch_size)

[@cyberfalcon2872]

@@sharkeys I'm on a 6gb 1660 btw

[@TINTO_BRO]

I don't understand this concept at all. I do not have a weak video card, but doing everything exactly as in the video I do not have enough memory, and when I lowered the value the number of steps is still prohibitive. So what values max_steps batch_size should I put?

[@sharkeys]

@@TINTO_BRO I'd suggest simply putting anything that you know is way to much ... I say this because you can always discard overbaked trainings and restart from the moment before it goes too far. Wait till the model overtrains then stop the training go grab a checkpoint from a few previous momments ago and then proceed to train that for a large amount of steps this time at a much lower learning rate... basically put an extra 0 in the learning rate, 0.005 goes to 0.0005 (for instance) I would personally depending on your amount of images go for two and split the stack. so for 10 images it would be Batch 5 Gradient 2

[@Thozi1976]

Great Video as usual. The textual Inversion Part (00:07:00-00:09:00) could have used a 2nd example, since I think it's key to get this right and it's a little confusing still after just one example given. Now I have to close the sandwich to accommodate the praise-criticism-praise rule of "suggestions for improvement". So keep it up - you're the King. 90% of what I know about AI Art I learned here in this channel. 🖖

[@YoZentrus]

Great comment, I enjoyed the humor of this "criticism sandwich" concept. However, it may be undermined somewhat when you introduce the possibility that the bread is only there to contain the meat & cheese. Open-faced sandwich.

[@endymion3213]

Any plans to make a new version of your DREAMBOOTH LOCAL Training tutorial? the new updates changed the settings menu and is confusing.

[@Aitrepreneur]

yes it's coming ;)

[@endymion3213]

@@Aitrepreneur

[@emillarsson271]

Has anyone managed to get this working on a mac? Python crashed as I hit "Train Embedding" and I also get the error "There appear to be 1 leaked semaphore objects to clean up at shutdown"...

[@stephenmilazzo2535]

Thanks. This video has been the best one for me and the results prove it. Excellent. Keep up the great work.

[@sub-jec-tiv]

Hi buddy! Can you explain how the process is different when training a style not a subject? Thanks this video is pure gold, i tried making an inversion but it looked terrible, i needed the info in this video!

[@Aleksandrsvideo]

I have GTX 1070 and when click Train Embedding i have next error: RuntimeError: CUDA out of memory. Tried to allocate 300.00 MiB (GPU 0; 8.00 GiB total capacity; 7.02 GiB already allocated; 0 bytes free; 7.22 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

[@RapPeriscope]

me too, write me when you have it done.

[@alross10]

Same error. No Idea how to fix it.

[@Grifter]

I was able to get to work by adding "--api --xformers --opt-split-attention" to the args in the webui.bat i have a 1070ti

[@CoqueTornado]

@@Grifter I do not know what I did but placing these args the python crashed saying "pandas" lacked some library blablabla... needed to reinstall python and automatic 1111 again... now solved... maybe I wrote something wrong :\

[@TizetGamer]

@@Grifter i have 1070ti. but how it work? can you let me know

[@bradcasper4823]

I have problem with saving embeddings. I'm using the same settings as on video. The problem is that every 50 steps training stops I have to manually click Train embedding in order to reach 3000 steps. In other words every 50 steps training stops and to continue it I have to click "Train Embedding" button again and again... Any help?

[@bradcasper4823]

Had to uncheck "Save images with embedding in PNG chunks".

[@helge666]

@Aitrepeneur I was wondering, if you omit the desired properties of a subject in the text ("long hair"), shouldn't you also omit "a woman", because Wednesday Addams is always a woman, and write "a person" instead?

[@Brian-Cherry]

Any help out there? I'm fairly sure I followed the instructions correctly. When I run it I get a message saying it finished in 0 steps. Anyone got any ideas? Thanks

[@marcelodesaellwanger4287]

same with me

[@nikfarmer]

run txt2img once with any prompt and then try to train embedding, worked for me.

[@Brian-Cherry]

​@nikfarmer thanks I'll try again.

[@olehlozovyi1605]

the best one lesson yet on the youtube. now i need new GPU to try this out

[@jgodvliet]

I followed are your steps and get the error: "CUDA out of memory. Tried to allocate 1.50 GiB". I am using a RTX 2080 TI with 11GB VRAM , 12 training images with batch size: 4 and Gradient accumulation steps: 3. Any idea how to get this to work or is my RTX 2080TI too old for this? Hope to hear from you.

[@jgodvliet]

I found the solution :) I removed the "--no-half in the line "set COMMANDLINE_ARGS= --no-half" from the web-user.bat

[@ZachFBStudios]

Mine has been working on 8gb with the following flags in the user bat: --api --xformers --opt-split-attention

[@Grifter]

@@ZachFBStudios It worked! i have a 1070ti 8gb vram and used those args to launch i did it with batch size 4 and gradient 6 and im using 24 images. Im curious what have you been able to get the batch size up to successfully with ur 8gb card? Update: I got it working all the way up to batch size 10 and gradient 2 using a dataset of 24 images and using the variable training in the video. But im still curious what you or other people are running on a 8gb card.

[@jgodvliet]

I used 12 images , 3000 max steps with batch size 4 and gradient 3. It took longer than expected 2.5 hours. 🤔 How long did it take you?

[@ZachFBStudios]

@@Grifter it looks like batch size 11 and gradient steps 4 is where i capped out. got cuda error at 12x4. I got it working on 9 x 5 and 7 x 3, and tried 10x1 based on your run without issue. my dataset is 45 images

[@RYOKUBE]

Please help me, getting stuck at 6:55... Preprocess cannot be proceed.. there's no txt files at my destination folder WHat should i do?

[@mauvaissigne]

i have the same problem. did you ever figure it out

[@DaKussh]

Pro tip: Always save your training resource in some place safe, specially things hard to get or that are tedious to make like processed images and their caption files, at some point people and project in general will start to move to newer versions of SD or any other popular model. At some point you'll have to update you own obsolete models for compatibility reasons.

[@tag_of_frank]

Could you please explain more advanced pre-processing, such as using GIMP to make the background transparent or putting some mask around the subject.

[@cdeford]

This explained some things that I found confusing in creating embeddings. How does it compare with dreambooth? Embedddngs are better than creating a large checkpoint file if the quality is as good.

[@30MinsGaming]

I have followed all the steps religiously and i get a 'Training finished at 0 steps' message after the training, i have tried to change the embedding learning rate from the formula to 0.000005 with the same results, can you please point to me where could be the problem?

[@30MinsGaming]

Details in case relevant: Running on a GTX 1080ti @ i7-8700K with 32Gb of RAM - Model selected is v1-5-pruned-emaonly.ckpt

[@30MinsGaming]

After some fiddling, I have found that setting both Batch size and Gradient accumulation steps to 2 allows the process to run, with the ETA of 2h for 20 images. Probably could have aimed a bit higher (?) but that were the first numbers allowing me to continue. Hope this helps someone.

[@abdullahkeles57]

Perfect, absolutely one of the most useful video for SD so far. Thank you :)

[@shalak001]

With the "use cross attention optimization while training" enabled, the embedding doesn't appear to be trained at all - the sample images generated are nothing like the training dataset. Once I disable the cross attention optimization, it trains correctly, albeit it forces me to use smaller batches (due to OOM errors) and the training time is quite big. Any idea why am I having issues with this option enabled?

[@dronematic6959]

for anyone who faced the similiar issue of finishing at step 0 with torch 2.0.0 and cuda 11.8 under the hood of rtx4090, upgrade to torch 2.0.1 with cuda 11.8 will solve the error.

[@sethlawson8544]

What GPU are you working with that you can do 11 images per batch? I'm getting a cuda out of memory error for every value except for 1 per batch 😭 8gb RTX 3060ti

[@objasen29]

My stable diffusion does not has the Preprocess images Tab on Training, anybody know why?

[@williamthebaud674]

Same here, and from what i found, since the 1.7 version of SD they deplaced it in the "Extra tab" but it's not that easy becose even if some options of preprocess images seems to be in the extra tabs, it hard to follow tutorial since the process just don't work anymore...

[@mauvaissigne]

@@williamthebaud674 i dont even see it in the extras tab. io am running veriosn 1.8.0 any idea on where i can find it?

[@valhalla0112]

Hello all, i hav a question, if my pc using nvidia geforce gtx 1050 ti with 4 GB , can i train this? what parameters should i use? i've been change the learning rate, batch, gradient, max steps, but it all did't work, it always says cuda out of mem. please help me. thx a lot 😃😃😇😇

[@hagwe]

Does it work well if you mix your "Textual Inversion" with others? For example, to give a certain style to the character?

[@sharkeys]

yes but you might need to play with the weight of the prompt using (brackets:0.6)

[@joshlake1882]

This is the one I've been waiting for, thank you!

[@Vigilence]

Would like to see this again, but covering art styles.

[@Epulsenow]

Hello Sir, can you help with my query about Hand and fingers deform ho when creating artwork like extra legs extra fingers. sometime the went missing or not corrects as per human anatomy. I tried negative prompts, but still issue remains.

[@DATRecords604]

thanks for this tutorial and everything else you have shared, thanks to your videos I have a locally installed stable diffusion and after hours of watching your content and trial and error, I am now capable of generating high-resolution images of great use. From this particular tutorial, however, I am not sure if I did everything right because all my results are very random images vaguely connected to the 70 I used to train the inversion. And on this note, I am wondering, if I were to use higher resolution images (let's say 768x768) do I need to use a 2.1 model? Or can I work with said resolution also with v1.5 (newbie question I know)

[@PeteStueve]

if planning to use this on dreambooth trained model would we then want to change woman to "wednesday" or whatever our instance prompt was ?

[@jeb123]

I tried with a RTX 3060 mobile and because of 6 GB vram i had to set batch size to 1 (i have 21 images). Processing time is 13 hours. I can't train that many hours unfortunately.

[@ilyakostin5676]

there must be some other way... RTX 3070 ti 8gb and ETA is 60-70 hours .-.

[@jeb123]

@@ilyakostin5676 I really gave up

[@autonomousreviews2521]

What a fantastic share! Great results. Thank you for your time :)

[@duonghuyhoang3940]

Can someone please tell me what model did Aitre use at 9:21? The 1-5 pruned one. I can't find it.

[@NicoTheVideoMaker]

hey! i'm currently trying to make an embedding but the results for now have been quite terrible lol Quick question: could it be because my preprocessed images DON'T get renamed with the content of the caption file? And in any case, is there a way to solve this?

[@tomalapapa100]

This is a great video! Congrats. From a point of view of a begginer with some ai coding experience it was great to get some technical stuf. Well done.

[@violainerudyjorisjacques7434]

Hello, Very good tutorial, very detailed! However, I'm stuck on one detail: Let's say we're going to create an integration with the suggested progressive learning rates. At 19:18 in the video, you indicate to resume training step 2750 with an embedding learning rate of 0.004. But in the progressive learning rates, step 2750 already trains at 0.001. Shouldn't we resume training this step at a higher level, maybe 0.0005? Or is there a subtlety in the formation of the integration that I haven't understood? If you could enlighten me!

[@elpioterro]

I thought the same thing

[@marcthenarc868]

Great video. I selected a male celebrity with vanilla 1.5. I trained it and did the grid but none of the tests resembled a photo, unlike your tests. They all looked ... OK but as if they were models from the old Poser program with harsh light and lots of wrinkles. I then I used my embedding with ProtogenX34 Photorealism and it came out pretty good, I then varied models with realistic, sketches, fantasy, heroic, etc. with great success. I can't explain the video game look at the training level.

[@ghostsquadme]

At 8:12 you describe taking out words that _directly_ relate to the character, wouldn't you remove "woman"?

[@anfanger4]

exactly my thoughts...

[@shadowdemonaer]

Bruh, I finally get it I think. Third time trying to get it and I'm finally having something click... But I need this for another reason. I am trying to train for negative prompts like how EasyNegative and a few others did. I am having some serious anatomical issues that should NOT be happening, and I would like to do something to correct them. with EasyNegative, you can just type that into the negative prompts area and it captures everything instead of having to do the usual "bad anatomy, worst quality, low res, bad hands, extra fingers" and so on and so forth. So does that mean if I train one for something, I could caption every image with the same short prompt like EasyNegative does, or is there more to it than this?

[@MrJohnSnow]

Thanks for the tutorial. It's very informative and easy to follow. I have several questions re: preselecting the images and re: editing the txt files with the descriptions: I want to make embeddings of all the main characters of the series "Buffy the Vampire Slayer"; starting with Willow Rosenberg, portrayed by Alyson Hannigan. Willow has a diverse but very "unique" fashion style. In order to capture this too, can I implement full and half body images of her in her typical clothes, or will this mess up the training of her face? What should be the percentage of face, half body and full body images? What is the maximum number of images before it becomes ridiculous or even counterproductive? Do the images have to show her standing upright, or can she be sitting or lying down, crouching, bending over, wildly gesticulating, etc, as well? Does it make sense to include some half and full body images that show her from the side or even from behind? When I generate pictures, I want her to be shown in Willow-typical clothes, unless I clearly specify in the prompt that she should wear something else (e.g. "picture of Willow wearing Navy Seal battle gear"). Do I remove the clothing descriptions from the txt files for that or do I leave them in? Lastly, it didn't come up in your video because Wednesday doesn't really show a range of facial expressions. Are those included in the txt files (sad, happy, angry, rage, crying, smiling, frowning, scowling etc.)? Thanks in advance to anyone who cares. :-)

[@TheCrazyCartModChannel]

this is the exact question I want to know, but no one has answered it as far as I can tell!

[@NONAME-ph1op]

For some reason I get an error at 7:01. TypeError: 'NoneType' object is not subscriptable. Do you know why? ? ?

[@RikkTheGaijin]

When I start the training I get this error: RuntimeError: CUDA error: invalid argument. CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1.

[@teebu]

same

[@markdavidalcampado7784]

same here

[@franzkekko2387]

Guys anyone knows why the textual_inversion folder doesn’t create itself to me? I can’t create my own embeddings because of this. I’m getting out of my mind 😭😭😭🙏🙏🙏

[@NiteSaiya]

I only had the VRAM to do a batch size of 5 and gradient accumulation of 4. 3000 steps took 6 hours, copies saved every 500 steps show no distinction from each other, and the output doesn't resemble my input at all. :/

[@Antares2]

I would recommend checking out Lora. There are several tutorials for using the kohya-ss GUI to do it. Training a face with around 20 images and 100-150 steps per image takes about ten minutes on my RTX 3070 and some of my Lora trainings have been very successful, giving a highly recognizable face.

[@ekke7995]

Im getting: RuntimeError: CUDA error: invalid argument CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. dose this mean my 12gb rtx3060 is to small or something ells?

[@edmartincombemorel]

yeah i had the same on my RTX 3090, remove the --medvram argument in your webui.bat file, worked for me

[@ekke7995]

lol, maybe this answers my own question; "Most of the time CUDA Runtime Errors can be the cause of some index mismatching so like you tried to train a network with 10 output nodes on a dataset with 15 labels. And the thing with this CUDA error is once you get this error once, you will recieve it for every operation you do with torch.tensors. This forces you to restart your notebook. I suggest you restart your notebook, get a more accurate trace back by moving to CPU, and check the rest of your code especially if you train a model on set of targets somewhere." I'll restart and try again now

[@ekke7995]

@@edmartincombemorel thanks, Im just restarting and will see again now

[@ekke7995]

I got it working; inside the webui-user.bat set COMMANDLINE_ARGS= ************************************* I removed --xformers --no-half basically left it empty.

[@edmartincombemorel]

@@ekke7995 yeah i did just that also, seems to work, no idea why but won't complain tho XD

[@b4st137]

Thanks a lot man! Great tutorial - that's exactly what i was searching for since weeks.

[@aknaysan6666]

can we do the same for any product? for example, an armchair or a sofa? The seat will remain stationary and will only change the ambience with the text prompt. Can this happen?