Software Engineering Courses (SECourses)•3y ago

SDXL quality is insane

mayna•10/7/23, 9:40 PM

how did you do that ? i want to know

kohya SS with SDLX lora training ?

Sandstormleader•10/7/23, 9:47 PM

i'm still pretty happy with 1.5

mayna•10/7/23, 10:36 PM

when you train , it's better to use sd1.5 or SD 2.1

SStefanoid9343 @Dr. Furkan Gözükara hi, I trying to do a training in dreamboth using runpod wh...

Furkan Gözükara SECourses•10/8/23, 12:32 AM

show screenshot of your workspace please

DDeleted User SDXL quality is insane

Furkan Gözükara SECourses•10/8/23, 12:33 AM

100%

SSandstormleader i'm still pretty happy with 1.5

Furkan Gözükara SECourses•10/8/23, 12:33 AM

super cool

Mmayna when you train , it's better to use sd1.5 or SD 2.1

Furkan Gözükara SECourses•10/8/23, 12:33 AM

sd 2.1 useless dont waste time

Dan•10/8/23, 1:02 AM

My googling is failing me: should I use the "Game Ready" or "Studio" NVidia driver?

Dan•10/8/23, 1:02 AM

I'm currently using the Studio Driver (I work a lot with DaVinci Resolve, so that made most sense to me at the time).

Dan•10/8/23, 1:03 AM

I'm having some system stability issues with loads of crashes, so I'm going through all options with reinstalling Windows last.

Stefanoid9343•10/8/23, 2:48 AM

Hi @Dr. Furkan Gözükara I follow your runpod tutorial for creating portrait photo images training model of a data set of a woman and the output start to create naked pictures half body there is any way to avoid that?

Sandstormleader•10/8/23, 3:25 AM

@Stefanoid9343 negative prompts, using models that do not have nudity, and i think there might be something in settings tab that helps as well.

Sandstormleader•10/8/23, 3:25 AM

Anyone have advice on best upscaling techniques?

Dan•10/8/23, 4:16 AM

I'm no expert, but I was spending a few hours the other night researching and experimenting with upscaling. I found this Reddit post that is really good: https://www.reddit.com/r/StableDiffusion/comments/y2mrc2/the_definitive_comparison_to_upscalers/

From the StableDiffusion community on Reddit: The DEFINITIVE Compar...

Explore this post and more from the StableDiffusion community

Dan•10/8/23, 4:17 AM

My favourite is still ESRGAN_4x, but I'm still trying new ones.

elie•10/8/23, 4:50 AM

Hi, could someone plz point me to the right resources for incorporating two safetensors into sd 1.5 or sdxl? Is it possible to use one that's completely different from the other (one that tells the model what my dog looks like & the other that is a pixel art theme)

Rrafraf i got pretty impressive results with batch 2, which drove SDXL adamW8bit trainin...

Khang•10/8/23, 5:28 AM

Very useful & interesting thank you very much. How is your flexibility test going? CivitAI also use batch 2 to train their Lora so I guess batch 2 could be a sweet spot.

virginia•10/8/23, 5:36 AM

sdxl is terrible on some notebooks

SSandstormleader @Stefanoid9343 negative prompts, using models that do not have nudity, and i thi...

Stefanoid9343•10/8/23, 5:53 AM

should I add negative prompts before start the training? also I’m getting double persons and half bodies instead of portrait shots. I’m training 768x1024 maybe is better use 1024x1024?

Khang•10/8/23, 6:01 AM

Hi guys how do you save Prompt as a text file with ComfyUI please?
I found this node of pythongosssss but I'm not a coder so I don't know how to change the "$output/*/.txt" to save the txt file into /workspace/ComfyUI/output/ComfyUI_00471.txt (next to the ComfyUI_00471.png so I can download both of them. I'm using Runpod)
https://github.com/pythongosssss/ComfyUI-Custom-Scripts

2023-10-08_12_49_10-text_file_di..._-_JupyterLab.png

2023-10-08_12_48_59-pythongosssss_ComfyUI-Custom-Scripts__Enhancements__experiments_for_ComfyUI_mo.png

Metaphysix•10/8/23, 6:11 AM

Anyone knows how to get kohya recognize the --train_text_encoder command?

Elisoli•10/8/23, 7:09 AM

do we have a config file or something to help us use kohya to train dreambooth on SD1.5 models?

virginia•10/8/23, 9:27 AM

anyone uses invoke ai, too?

virginia•10/8/23, 9:27 AM

why it seems that each prompt can produce only 1 image at one time

feritkaya34•10/8/23, 9:36 AM

Initializing bucket counter!
Steps: 1%| | 10/1100 [01:38<1:59:03, 6.55s/it, inst_loss=0.0501, loss=0.0501, lr=0.0001, prior_loss=0, vram=6.8] guys is this a normal ? I started use brand new stable diffusion v1.6. If you got advice how can i increase this speed please share me. Sorry for my bed england

feritkaya34•10/8/23, 9:37 AM

i got rtx 4070 8gb r7 7840 hs cpu

Sandstormleader•10/8/23, 10:13 AM

@Stefanoid9343 from my experience the double the double people thing really depends on the model you are using to train with because if it was trained in 512 or you are much more likely to get double people if you go higher than 768 and it's still not always perfect if it was a 768 training. a lot of the newer really good models i have downloaded have specified that you need to specify outfits or you are more likely to get nsfw results. your best bet if you don't want any chance of nude images is to find a model that doesn't produce nsfw images i think. but i am no expert.

LEGACIES•10/8/23, 10:15 AM

does anyone know what is this error?

LEGACIES•10/8/23, 10:37 AM

@Dr. Furkan Gözükara I am using 150 training steps per image and have 30 images in my dataset. I am not using any classification images. Do you have any advice why does my train model look different from my dataset?

LEGACIES•10/8/23, 10:37 AM

Sandstormleader•10/8/23, 10:44 AM

what tokens are you using?

Sandstormleader•10/8/23, 10:45 AM

if you just gave it a standard name the model could already have some info on that for example if you are using "olivia" vs something like "ohwx"

LEGACIES•10/8/23, 10:57 AM

Sandstormleader•10/8/23, 10:57 AM

i don't even put woman in my instance prompt personally

Sandstormleader•10/8/23, 10:59 AM

not sure if that matters but @Dr. Furkan Gözükara usually suggest class images when he does it that way or when he trains in general

LEGACIES•10/8/23, 11:01 AM

i will try to train another one without the word woman

LEGACIES•10/8/23, 11:02 AM

how many class images do i need for 30images in dataset?

Sandstormleader•10/8/23, 11:03 AM

that should be plenty from my experience

Sandstormleader•10/8/23, 11:04 AM

i've done some with 20 and others with 200

Sandstormleader•10/8/23, 11:05 AM

over training and bad quality have usually been my only issues with not using class images, it's pretty much always looked like my dataset

LEGACIES•10/8/23, 11:09 AM

so you always used class images when doing training?

Sandstormleader•10/8/23, 11:09 AM

not me no,

Sandstormleader•10/8/23, 11:09 AM

i also don't downsize my images or anything like that

Sandstormleader•10/8/23, 11:10 AM

i read somewhere on reddit that it actually does all that for you in theory. idk anything official but i know it's been working for me

LEGACIES•10/8/23, 11:10 AM

training model is so difficult XD

Sandstormleader•10/8/23, 11:10 AM

are you using 1.5 or sdxl?

LEGACIES•10/8/23, 11:11 AM

1.5

Sandstormleader•10/8/23, 11:11 AM

is it a lora or ckpt?

LEGACIES•10/8/23, 11:11 AM

checkpoint

Sandstormleader•10/8/23, 11:11 AM

ok ya that's same as me

SDXL quality is insane

Similar Threads