Software Engineering Courses (SECourses)•3y ago

Need some help, need to ideate on what the background should look like. Is there something in contro

Need some help, need to ideate on what the background should look like. Is there something in control net I can use to ideate some background ideas?

Cc: @Dr. Furkan Gözükara

Darth Teethius•12/27/23, 5:34 AM

Anyone recommend a good lora dim/alpha for realistic sdxl models? Is 128/1 still best or has everyone moved to the dreamboat extract mode now?

FFurkan Gözükara SECourses usually 15 images with 150 repeat yields good results even little bit overtraine...

Baran•12/27/23, 11:10 AM

for one epoch?

BBaran for one epoch?

Zax•12/27/23, 11:39 AM

if I'm correctly understanding what he's saying, an "epoch" is a "full run" of training on your data set. So multiple epochs constitutes multiple runs of training on your data set. This is functionally the same thing as doing repeats for images. so your total number of steps is [training images] x [repeats] x [epochs], and take that and double it if you have regularization images. So doing 15 training images with 150 repeats and 1 epoch is 2250 steps (or 4500 with reg images). If you instead did 15 training images with 10 repeats for 15 epochs, that would also be 2250 steps (4500 with regs).

Max•12/27/23, 12:48 PM

guys I have a question. Why does it give me a different output every time I copy a prompt from civitai and paste it into sdxl with the same settings? Shouldn't he give the same? Do I have to install like texture inversion? Help me

Lorelai AI•12/27/23, 1:21 PM

@Dr. Furkan Gözükara I am playing around with different settings and one model is giving me issues in training a Laura. Most go smooth based on your video. I've tried more pictures, less pictures, more repeats, etc. I'm trying a higher learning rate right now. My go to settings are usually 2 batches(4090 doesn't appear to handle 3 or more on AdamW) 15 repeats and 15 Epochs. Default learning rate. This has worked awesome with 14 images or I've had great results with 8 repeats, 15 epochs, default learning rates, but 29 images. My subject is a weightlifter and I'm training with SDXL Base. Any tips?

Ppaleolithic Need some help, need to ideate on what the background should look like. Is there...

Furkan Gözükara SECourses•12/27/23, 1:56 PM

what are you trying right now?

AAIGambino Click to see attachment

Furkan Gözükara SECourses•12/27/23, 1:57 PM

something else is looking wrong

Furkan Gözükara SECourses•12/27/23, 1:57 PM

not using vram

DDarth Teethius Anyone recommend a good lora dim/alpha for realistic sdxl models? Is 128/1 still...

Furkan Gözükara SECourses•12/27/23, 1:59 PM

it need further testing

Furkan Gözükara SECourses•12/27/23, 1:59 PM

but i tested lora extract

Furkan Gözükara SECourses•12/27/23, 1:59 PM

128x128 and 192x192 and 256x256 almost same

Furkan Gözükara SECourses•12/27/23, 1:59 PM

when you extract

Furkan Gözükara SECourses•12/27/23, 1:59 PM

out of dreambooth model

BBaran for one epoch?

Furkan Gözükara SECourses•12/27/23, 1:59 PM

yes. because 150 repeating making it 150 epoch actually

Furkan Gözükara SECourses•12/27/23, 1:59 PM

since every training image is 150 times trained

MMax guys I have a question. Why does it give me a different output every time I copy...

Furkan Gözükara SECourses•12/27/23, 2:14 PM

seed matters

FFurkan Gözükara SECourses seed matters

Max•12/27/23, 2:15 PM

but i put them

LLorelai AI @Dr. Furkan Gözükara I am playing around with different settings and one model i...

Furkan Gözükara SECourses•12/27/23, 2:16 PM

i noticed same. sdxl realistic vision was undertrained at my settings

MMax but i put them

Furkan Gözükara SECourses•12/27/23, 2:16 PM

could have extra loras etc

FFurkan Gözükara SECourses i noticed same. sdxl realistic vision was undertrained at my settings

Lorelai AI•12/27/23, 2:50 PM

I used a Learning rate of .001, thew results were much better! .001 might be high but because I'm doing parody I don't really need exact I just need to capture their likeness and make them something else

Enlil Wolf•12/27/23, 3:23 PM

@Dr. Furkan Gözükara - Dr, when you get the chance, please check the DM

LLorelai AI I used a Learning rate of .001, thew results were much better! .001 might be hig...

Furkan Gözükara SECourses•12/27/23, 3:24 PM

ye it changes according to settings

B0nd•12/27/23, 4:35 PM

When training for SDXL DreamBooth or Lora in KoyhaSS what is generally the best way, for likeness and flexibility?

With or without regularization images?
With or without captioning the dataset?

FFurkan Gözükara SECourses you need to use our 16gb config

DigitalArt_Lover•12/27/23, 5:41 PM

I need to know, the 16gb config, can i get it from you by any chance?? ..i would buy the Patreon way long ago, but for my country i could pay, because of the payment method restrictions.

BB0nd When training for SDXL DreamBooth or Lora in KoyhaSS what is generally the best ...

Furkan Gözükara SECourses•12/27/23, 7:54 PM

it depends purpose. i made a good comparison here recently

Furkan Gözükara SECourses•12/27/23, 7:55 PM

https://civitai.com/articles/3510/dreambooth-training-on-sdxl-anime-model-and-comparing-ground-truth-regularization-images-effect

DreamBooth Training on SDXL Anime Model & Comparing Ground Truth Re...

It was requested of me to test ground truth Regularization / Classification images effect during Stable Diffusion XL (SDXL) DreamBooth training. So...

DDigitalArt_Lover I need to know, the 16gb config, can i get it from you by any chance?? ..i would...

Furkan Gözükara SECourses•12/27/23, 7:55 PM

hello. sure i accept crypto and other methods too

Furkan Gözükara SECourses•12/27/23, 7:55 PM

you can dm me

Benjamin•12/27/23, 8:14 PM

@Dr. Furkan Gözükara did I get it right, you are now training with 150 repeats, 1 epoch for 13 training images and save steps instead of epochs?

Benjamin•12/27/23, 8:15 PM

so 150 * 13 * 1 *2 = 3900 steps?

Digital [Starburst]•12/27/23, 8:21 PM

I loaded the styles.csv into Automatic and have gotten some very interesting results using the updated Juggernaut 7 model. The advantage to this model is the refiner is built into the model, so you don’t need a separate refiner. The disadvantage to using the styles in Automatic is that you cannot combine styles in one image. If you select multiple styles you will get a batch for each style you choose.

Digital [Starburst]•12/27/23, 8:24 PM

In the Fooocus UI you can combine styles for more variety in your art. You can also use the same model by manually adding it. I personally like to select one style and generate a prompt to see the results before adding another style or switching styles. The Fooocus Masterpiece style is my favorite so far.

Digital [Starburst]•12/27/23, 8:26 PM

Combining two styles is, in essence, a completely different style. As you add styles to a prompt you get unexpected results.

Digital [Starburst]•12/27/23, 8:28 PM

The caveat to combining styles is one may completely cancel out others.

Digital [Starburst]•12/27/23, 8:33 PM

I take my images, load them into Corel Painter or Corel Paintshop Pro and adjust them, make them fit an 8.5 X 11 sheet of paper or 11 X 17 sheet for printing. I also may completely alter the image because I don’t want photographs even though I want to study photography in order to write good prompts. At the moment I borrow other people’s prompts and modify them. I don’t want to be reliant on someone else’s prompts forever though.

Digital [Starburst]•12/27/23, 8:35 PM

It seems that the general consensus for most AI developers is they want photographic art. That’s all well and good but they are really limiting themselves in my opinion as SD is capable of so much more.

Digital [Starburst]•12/27/23, 8:38 PM

True. Not without the use of other digital art tools but the results I am seeing can’t be ignored.

Digital [Starburst]•12/27/23, 8:39 PM

Digital [Starburst]•12/27/23, 8:42 PM

That is a raw image straight from Fooocus UI using 3 styles but the prompt itself also specifies a style which overrides apparently the selected styles.

allen•12/27/23, 10:37 PM

Howdy all - I'm trying to get MagicAnimate working in Runpod, can anyone point me to a thread or discussion?

allen•12/27/23, 10:38 PM

I tried it in comfyUI local install and it just seems buggy, the nodes are anyway, and then I realized the good Dr had already put together a better implementation. I have the Gradio space working right now in a runpod

lajotas.•12/27/23, 10:43 PM

@Dr. Furkan Gözükara , when training a SDXL model with 30 training images of a subject for a person's resemblance, after how many epochs do you expect the images to start looking like the person?

BBenjamin @Dr. Furkan Gözükara did I get it right, you are now training with 150 repeats, ...

Furkan Gözükara SECourses•12/28/23, 12:00 AM

yes