Software Engineering Courses (SECourses)•3y ago

Hi guys. I am struggling LoRA training in Dreambooth on Automatic1111. If I train a normal Dreamboot

Hi guys. I am struggling LoRA training in Dreambooth on Automatic1111. If I train a normal Dreambooth I get fairly good images. When I switch to LoRA the samples are okay but when I try a prompt with the LoRA it gives me completely random images. Is the extension broken or is it likely user error?

AlmoIppy•3/11/23, 11:02 PM

I was under impression that with dreambooth I could make a large dataset (say 1000 images or more) and create individual prompts for each image that describes things from it I want to teach the model.

AAlmoIppy I was under impression that with dreambooth I could make a large dataset (say 10...

Furkan Gözükara SECourses•3/11/23, 11:02 PM

correct

Furkan Gözükara SECourses•3/11/23, 11:02 PM

all you need is good images and captions

MMajor Epidemic Hi guys. I am struggling LoRA training in Dreambooth on Automatic1111. If I trai...

Furkan Gözükara SECourses•3/11/23, 11:02 PM

lora is much harder

AlmoIppy•3/11/23, 11:02 PM

Yeah but in the extension there is single field for a prompt.

Furkan Gözükara SECourses•3/11/23, 11:02 PM

requires fine tuning of parameters

AAlmoIppy Yeah but in the extension there is single field for a prompt.

Furkan Gözükara SECourses•3/11/23, 11:02 PM

watch this i explain logic of filewords

Furkan Gözükara SECourses•3/11/23, 11:03 PM

https://youtu.be/KwxNcGhHuLY

YouTubeSECourses - Software Engineering Courses

DreamBooth Got Buffed - 22 January Update - Much Better Success Tra...

Our Discord : https://discord.gg/HbqgGaZVmr. Newest update of DreamBooth extension of Automatic1111 brought huge quality and success improvement. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of Stable Diffusion Tutorials...

AlmoIppy•3/11/23, 11:03 PM

Thanks.

Furkan Gözükara SECourses•3/11/23, 11:03 PM

i sent a dm to you if you be interested in

Facundo Túpac•3/11/23, 11:47 PM

Howdy! First of all I want to thank for all the great content provided in the YouTube channel. I've been watching a few to better understand the subject but I wonder which specific video or videos I should watch which is the most up to date if what I want to do is to use Stable Difussion with my own face with the best quality possible in Google Collab. I'd also like to do the same with a 2nd person. Should I train a 2nd model from zero or could I add the 2nd person to an already trained model? If so, how to do it for the best quality possible?

Facundo Túpac•3/11/23, 11:47 PM

I don't currently have an Nvidia card good enough for local

FFacundo Túpac Howdy! First of all I want to thank for all the great content provided in the Yo...

Furkan Gözükara SECourses•3/11/23, 11:58 PM

ok i presume you cant pay to use runpod right?

Furkan Gözükara SECourses•3/11/23, 11:58 PM

so now i will list some tutorials for you

Furkan Gözükara SECourses•3/11/23, 11:58 PM

this one how to train on colab

Furkan Gözükara SECourses•3/11/23, 11:58 PM

still works

Furkan Gözükara SECourses•3/11/23, 11:58 PM

https://youtu.be/mnCY8uM7E50

YouTubeSECourses - Software Engineering Courses

Transform Your Selfie into a Stunning AI Avatar with Stable Diffusi...

Our Discord : https://discord.gg/HbqgGaZVmr. How to do free Stable Diffusion DreamBooth training on Google Colab for free. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of Stable Diffusion Tutorials, Automatic1111 and Goo...

Furkan Gözükara SECourses•3/11/23, 11:59 PM

this is a quicker and more up to date version of colab

Furkan Gözükara SECourses•3/11/23, 11:59 PM

also it shows more info about how to use colab path

Furkan Gözükara SECourses•3/11/23, 11:59 PM

https://youtu.be/kIyqAdd_i10

YouTubeSECourses - Software Engineering Courses

Stable Diffusion Google Colab, Continue, Directory, Transfer, Clone...

Our Discord : https://discord.gg/HbqgGaZVmr. This is the video where you will learn how to use Google Colab for Stable Diffusion. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of Stable Diffusion Tutorials, #Automatic1111...

Furkan Gözükara SECourses•3/11/23, 11:59 PM

this is my newest dreambooth settings comparison but it uses web ui

Furkan Gözükara SECourses•3/11/23, 11:59 PM

https://youtu.be/sRdtVanSRl4

YouTubeSECourses - Software Engineering Courses

Epic Web UI DreamBooth Update - New Best Settings - 10 Stable Diffu...

RunPod: https://bit.ly/RunPodIO. Discord : https://bit.ly/SECoursesDiscord. Web UI DreamBooth got epic update and we tested all new features to find the best settings (10). If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of...

Furkan Gözükara SECourses•3/11/23, 11:59 PM

and this is a massive tutorial that explains concepts of dreambooth

Furkan Gözükara SECourses•3/11/23, 11:59 PM

https://youtu.be/Bdl-jWR3Ukc

YouTubeSECourses - Software Engineering Courses

Zero To Hero Stable Diffusion DreamBooth Tutorial By Using Automati...

Our Discord : https://discord.gg/HbqgGaZVmr. The most advanced tutorial of Stable Diffusion Dreambooth Training. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of Stable Diffusion Tutorials, Automatic1111 and Google Colab ...

Furkan Gözükara SECourses•3/12/23, 12:00 AM

teaching 2 faces into 1 dreambooth model is a bit problematic. if they are different gender can work but if same gender

Furkan Gözükara SECourses•3/12/23, 12:00 AM

then you should not use classification images to not bleed

junsama•3/12/23, 12:57 AM

Hi professor, have your trained lora with Kohya's webui

junsama•3/12/23, 12:57 AM

I see a lot of videos talking about that

junsama•3/12/23, 12:57 AM

I trained a lora with it and applied to the sd1.5

junsama•3/12/23, 12:58 AM

nothing happend

junsama•3/12/23, 12:58 AM

Kohya's lora also modifies the unet, which other loras from civitai does not have

FFurkan Gözükara SECourses this is impossible. it depends on your dataset

n8hawk_•3/12/23, 1:11 AM

what do you mean by "your dataset"? certainly it depends on the quality of your pictures. but as long as you have (like the pictures that you are used to take for your comparisions) simialar pictures you can compare f.e. 3, 6, 10, 20, 35, 50 pictures for training to find out how near the different trained images come to the "original" person..

Steel•3/12/23, 1:15 AM

the number of pictures really does not matter for "accuracy", but it matters for variety, you can train with one single image, and have that exact look on all your outputs

n8hawk_•3/12/23, 1:22 AM

ah, that makes sense. so the effects is similar to overtrainig as an overtrained image shows mainly the unchanged input images

Steel•3/12/23, 1:22 AM

you cant overtrain a lora

n8hawk_•3/12/23, 1:23 AM

ok- i did not refer to lora in my main question

Steel•3/12/23, 1:24 AM

oh that wasnt you, ok then.. well yes you can put something in your sanity text field that has nothing to do with your subject or like a specific style

Steel•3/12/23, 1:24 AM

and when that dissaplears from the samples, you overtrained

n8hawk_•3/12/23, 1:25 AM

yes.

Steel•3/12/23, 1:27 AM

I think there are several understandings of what overtraining is. you can overfit by destroying your subject. Like deform it. and you can overtrain by making it not flexible

Steel•3/12/23, 1:28 AM

but I wonder why people would train a model baked into a checkpoint, you by default make it non flexible

n8hawk_•3/12/23, 1:30 AM

i simply save at a alow epoch frequence so I can later compaare how good the single checkpoints are

Steel•3/12/23, 1:30 AM

thats not my point

Steel•3/12/23, 1:30 AM

you have a checkpoint thats good for one thing but not another

Steel•3/12/23, 1:31 AM

one is good for like steampunk or scifi, others are good for fantasy

Steel•3/12/23, 1:31 AM

none are good for everything

n8hawk_•3/12/23, 1:32 AM

i think in this respect the source image is more relevant

Steel•3/12/23, 1:32 AM

but you want a nice background dont you?

n8hawk_•3/12/23, 1:34 AM

that depends of the source image and the flexibilty of the trained image to "translate" your text input

Steel•3/12/23, 1:35 AM

Hi guys. I am struggling LoRA training in Dreambooth on Automatic1111. If I train a normal Dreamboot

Similar Threads