Dreambooth Face Training: Google Colab's SUPER EASY Method

Рет қаралды 14,326

Жыл бұрын

Stable Diffusion is AMAZING! .... However ....
Not every face you want has been trained. Using this free method you can train ANY face you like within 40 mins and have great results with just a little tweaking. All resourses used in the video are linked below. Enjoy!
- MY SETTINGS -
The Last Ben Colab
UNet_Training_Steps - 3000
UNet_Learning_Rate - 2e-6
Text_Encoder_Training_Steps - 250
Text_Encoder_Learning_Rate - 2e-6
Negative Prompts for Stable Diffusion (Works for other AI services)
disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, mangled, old, surreal, out of frame, ((((mutated hands and fingers)))), deformed eyes, ((bad art)), extra fingers, cloned face, body out of frame, gross proportions, canvas frame, cartoon, 3d, weird colors
- LINKS-
Automatic1111 (Local Stable Diffusion) -
github.com/AUTOMATIC1111/stab...
HuggingFace (For tokens and to accept the SD model cards)
huggingface.co/
TheLastBen Github
github.com/TheLastBen
TheLastBen Colab
colab.research.google.com/git...
Birme (For easy perfect resizing of instance images)
www.birme.net/
Replicate - GFPGAN (AI upscaling of finished image)
replicate.com/tencentarc/gfpgan
Gettyimages (High Quality face photos for PERSONAL USE)
www.gettyimages.co.uk/editori...

Пікірлер: 78

@dustboii Жыл бұрын

Okay I just came to say one thing... I have been scratching my head for 3 days at why my trained models look horrible no matter what settings i use. Come to find out It was because I was clicking restore face... It was only after watching your video I learned this.

@leogassie3378 Жыл бұрын

Thank you a lot !! Very useful video 🔥🔥

@satishpillaigamedev Жыл бұрын

cool workflow, thanks a lot

@flowerpixel Жыл бұрын

subscribed, i learned about SD and google colab 2 days ago and i just tried this as a complete newbie. it works

@ThunderMonique Жыл бұрын

thank you so much for this fantastic tutorial, DigitalDaz! I want to upload my class images but don't know how to do it. What should be included in the code?

@nimaafsari87 Жыл бұрын

Great video. I have a question. Would these same principles apply if your merging this kevin bacon model with another model? What % do you prefer the AI to favor kevin bacon (model a) vs the merging model (model b)?

@izakhatton588 Жыл бұрын

Great video, quick question, do you not need to specify an instance prompt or a class prompt for fast dreambooth?

@moibe182 Жыл бұрын

Wow! the renaming of files is a nice trick that probably have been there for years and I hasn't realized it existed, nice, thanks.

@solraksolrak Жыл бұрын

Muy bueno, muchas gracias.

@emrey2252 4 ай бұрын

awesome. thank you

@hartiym Жыл бұрын

What should I do in the Download Model section? When it doesn't, it gives the error "No model found, use the "Model Download" cell to download a model."

@iamstarman7 Жыл бұрын

Closeups and portraits the face looks really good, but "far away" images it's seems like a different face. Inpainting doesn't change the face in the way I would hope. Any thoughts on settings? Maybe new version changed things?

@rodrigomenam Жыл бұрын

Thank you so much. Appreciate this tutorial, I can finally see how I'd look GTA style :)

@Gh0sty.14 Жыл бұрын

Thanks so much for this! It's honestly amazing how well it's working for the face. Only issue I'm having is it's practically ignoring all other prompts, Location, objects, clothing, body position etc. Nothing I try seems to work well at all for that when using my custom model.

@Gh0sty.14 Жыл бұрын

@@DigitalDazYT Yeah I was thinking it was probably the pics I used. They all had a white background and were cropped to only contain the face. Oh well, time to make another!

@Gh0sty.14 Жыл бұрын

@@DigitalDazYT Another question.... is it possible to train on a different model? I'd like to use realistic vision v1.2 instead of the SD model.

@arafatboss6067 Жыл бұрын

Become your 500 one subscriber😁 good luck bro

@arjuns2020 Жыл бұрын

Hi bro, one more doubt, suppose if we are training multiple set of characters like 6-7 characters, each having 10 photos, how much steps do we need to keep ?

@LinkRammer 11 ай бұрын

Does this method work the same if I wanted to train a style? I want to make a liminal spaces model.

@vivalsdne3239 Жыл бұрын

hi , can i train model in google colab using custom model like protagen or dreamlike diffusion ? i dont want to use base stable diffusion v.1.5 model

@helsia9893 Жыл бұрын

Hey! How do the "external captions" work? Should I make a txt file with the same name as each image file and then write inside of the notepad the prompt/caption of the picture? And do I leave that text file in the same folder as the pictures?

@helsia9893 Жыл бұрын

@@DigitalDazYT Thank you so much for the help!

@hunterscruggs5316 Жыл бұрын

could you please explain how to get link for training on different models? I tried for liberty and URPM and neither link was valid. hat links am I missing?

@planetfrog2 Жыл бұрын

Great stuff, giving this a go. Would it work for objects ( like a dog ) too or is it specifically geared to faces?

@planetfrog2 Жыл бұрын

@@DigitalDazYT Thanks Daz, yeah I've been trying to work with Dreambooth on Automatic1111, but its slow progress, and I just feel like I'm not achieving much once I get a ckpt file out of it. Hopefully this will speed up the whole learnign process. Will check out the captions and regularization sections as you suggest. Thanks again :)

@planetfrog2 Жыл бұрын

@@DigitalDazYT Headsup....after the training and testing the CKPT - IT WORKED very well!!!! - thanks so much - your tutorial has helped me break through the barrier!

@neosyoshiranquake6160 Жыл бұрын

becasue resolution your images its 1024x 1024 and not 512x512? when PPI resolution?

@nimaafsari87 Жыл бұрын

Why do you have to get the token from Hugging Face in the first place?

@gowiththeflowguy8310 Жыл бұрын

Hello, im getting an error when i try use my ckpt file it says in the command prompt "The file might be malicious, so the program is not going to read it. You can skip this check with --disable-safe-unpickle commandline argumen"

@lifelens8995 Жыл бұрын

There are two issues, 1) If you don't specify any token or model, which model will it use bydefault ? 2) It gives an error wget not found even if all dependencies are installed.

@JTsWorks Жыл бұрын

For some reason my checkpoint won't show up on the dropdown. I have it in the models folder.

@Bartskol Жыл бұрын

Can it be done with 2.1 on colab?

@neosyoshiranquake6160 Жыл бұрын

In min 25 . Any option for resize local?

@ELSuperJake Жыл бұрын

I dont have the option to "inpaint at full resolution" when I go to inpaint

@tbo7163 Жыл бұрын

Thank you so much for a the video!! one question, i see on your video, on the img2img/ inpaint tab, there is an option called "inpaint at full resolution" Right but i can't find it on mine,please help!!!!!!!!!

@tbo7163 Жыл бұрын

@@DigitalDazYT Thank you for the response!!!! but i updated it today, i also i have "git pull" command for automatic update too. I think they removed it, can you check it on yours now? and let me know? Thank you again:)

@tbo7163 Жыл бұрын

@@DigitalDazYT oh okay i was worried thinking i had an outdated UI.. THANK YOU SO MUCH FOR THE INFO🙏🙏🙏 Great Content! i'll be waiting for your next video!!

@maxbrych7354 11 ай бұрын

do the image files must be in the same file format? or can I train with png's and jpg's?

@DigitalDazYT 11 ай бұрын

It helps if they are the same format. You can use online converters which are free or if you're using a PC there is free software called Format Factory that can do it too.

@barrilo Жыл бұрын

thank you very much for your video. I have a problem, I don't see the space to paste the hugginface token :(

@odawgthat3896 Жыл бұрын

@@DigitalDazYT thx

@ZainAli-hb4cm Жыл бұрын

This is one of the best explanation video. I was thinking about what else can we train the model with instead of people's faces. Can we train it to create a specific shoe or clothes on a model? please let me know how can i do that. Thanks, @digitaldaz

@zain1045 Жыл бұрын

Heya! Can I train models on colab with 4 gb vram using Dreambooth ? (1650s)

@zain1045 Жыл бұрын

@@DigitalDazYT ah i see thank you!

@hondas6463 Жыл бұрын

i only have Path_to_HuggingFace not HuggingFace toke what I do

@paveldesort5657 Жыл бұрын

this is the best tutorial I found to get colab to make model of a face with 15-20 instance images for SD 1.5. I followed all the settings to the letter and great results. However, I tried to replicate the steps for SD 2.1 768 model with 768x768 instance images of the same source and the face was not recognizable. I pushed the weight to 2.2 which somwhat helped but only marginal. Would you know how to train 2.1 768 model in this case? have you tried it? Again, cheers for this tutorial ;)

@paveldesort5657 Жыл бұрын

@@DigitalDazYT I have had just now a luck with weight of :1.5 on instance word (sd 2.1 768) using the ckpt from lastben's fast dreambooth notebook. What DIM number (Lora UNET Rank) does this fast dreambooth trains with? I was using DIM 128 number for LoRA extraction tool successfully on SD1.5 from this fast dreambooth but with 2.1 DIM 128 (Lora UNET Rank) doesnt work.

@chardy7071 Жыл бұрын

Do you have a tutorial for a newbie? (Totally new to this)

@chardy7071 Жыл бұрын

@@DigitalDazYT Probably the right question would be, which one should be install, invoke or Stable diffusion (which has a lot of steps)?

@arjuns2020 Жыл бұрын

Hello dude, is it possible to train our images in existing models like Deliberate/Realistic_vision ?

@arjuns2020 Жыл бұрын

@@DigitalDazYT Pasting model link from hugging face is not working apparently, will it work if I upload the model to gdrive, then paste that link on collab ?

@arjuns2020 Жыл бұрын

@@DigitalDazYT Thanks mate 🙂

@user-nx5yj4sk3t Жыл бұрын

I used 17 photos and 3000 steps but the results are terrible. How do I know if I have not trained the model enough, or it was overtrained? (i used your settings in gradio and also i tried to change sampling steps/cfg, it didn't help)

@user-nx5yj4sk3t Жыл бұрын

@@DigitalDazYT yes. I trained another model with checkpoints every 500 steps and the results disappointed me. Sometimes i can get good results, like 2k steps trained model and 125 steps with cfg 5 sd settings without restoring face, bit in most cases it looks bad. Looks similiar, but another human, maybe it's in the photo. I suppose a person should look exactly the same in all photos, which is almost impossible if we are talking about an ordinary person, and not about a popular actor/singer. I used a studio photoshoot, but in different styles and outfits and it turned out to be a mess.

@user-nx5yj4sk3t Жыл бұрын

@@DigitalDazYT i can send you x/y plot with different cfg/steps and some references, if you interested to see the result. Maybe a little more or less steps are needed sometimes? I just don't know how exactly to define MORE or LESS, are there any criteria?

@mohammedirfan6717 Жыл бұрын

@@user-nx5yj4sk3t any updates on your results?

@HighTechLogistics Жыл бұрын

It doesn't work, "check the link you provided" But there is no " accept button " for the license, Is there anyone able to help with that or maybe has fresh colab to use? Any kind of help will be highly appreciated, thanks

@vitor5184 Жыл бұрын

same problem did you solve it ?

@HighTechLogistics 5 ай бұрын

yes, don't insert link, but manually choose the way to the file@@vitor5184

@gilmarrodriques2113 Жыл бұрын

How to download this automatic 1111 I am not able to download,

@gilmarrodriques2113 Жыл бұрын

@@DigitalDazYT until the end of the page

@vitor5184 Жыл бұрын

It pops up a error can you help me its ''Check the link you provided'' while Path_to_HuggingFace

@edsonmatheus7976 Жыл бұрын

Same here

@vitor5184 Жыл бұрын

@@DigitalDazYT thank you very much, i will

@vitor5184 Жыл бұрын

@@edsonmatheus7976 just don't put anythink in the Path_to_HuggingFace should work fine

@faraday8280 Жыл бұрын

Help i trained a model but it does not show up at GDrive

@capasi5380 Жыл бұрын

@@DigitalDazYT btw which GPU you use on your computer?

@StoriesWithAPurpose Жыл бұрын

@@DigitalDazYT waited 5m, still not there. it did work before, not today

@faraday8280 Жыл бұрын

@@DigitalDazYT thx it was my input error lol

@heranzhou6976 Жыл бұрын

@@faraday8280 what error did you make? I also could not find mine in Google Drive

@user-cd4fb7nb9o 5 ай бұрын

just click model download on that step since is not need to get a hugging face token

@user-cd4fb7nb9o 5 ай бұрын

nevermind my ckpt model didnt even load in stable diffusion

@lenimann 3 ай бұрын

The part where i should enter the Hugging Face Token does not appear.

@innocentbystander6674 3 ай бұрын

You don't need to do it anymore. Just skip that step

@thatonegamer117 2 ай бұрын

@@innocentbystander6674it doesn’t work if you skip that step

@patrickhollywood93 Жыл бұрын

Love your work! Thanks for the tutorials! -🦾🎶🦋

@neosyoshiranquake6160 Жыл бұрын

for other model. i need tokken? ckpt_link? min 6:38