Software Engineering Courses (SECourses)•2y ago

<@205854764540362752> when you get some time can you try this prompt out on your training? Postive

@Furkan Gözükara SECourses when you get some time can you try this prompt out on your training?

Postive - cinematic photo Portrait of ohwx man as super Mario, red hat, fantasy, highly detailed, digital painting, artstation, concept art, sharp focus, illustration, art by Tony Sart and artgerm and randy vargas. 35mm photograph, film, bokeh, professional, 4k, highly detailed

Negative - drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, 3d render

DPM ++2sa Karras

1024x1024

CFG - 7

can you check if you get good likness

Théo•7/11/24, 2:09 PM

After 30 training and 5 dataset retake to have the best one, i finally achieve the - p e r f e c t - result thanks to @Furkan Gözükara SECourses however, mid-distance or distance shot are Totally destroyed, my dataset is perfect, used cinema grade camera, used cropper and resizer, tried mask training, adetailer, changing the number of step only for face, changing the denoising strengh, i can only make p r e f e c t close-up, selfie or whatever, but when the subject is in midrange like your picture in civtai, or fullbody, it's totally crap, even the face is not really "human" more like destroyed, any idea ? thanks for everything (also tried with and without reg)

TThéo After 30 training and 5 dataset retake to have the best one, i finally achieve t...

MaxiviOP•7/11/24, 2:11 PM

are you training on realvis4 and using the model on a different checkpoint?

Théo•7/11/24, 2:32 PM

traing on realvis4, using on realvis4

Théo•7/11/24, 2:33 PM

i have to admit i spent the last weeks understanding every little things possible and even myself cannot recognize a (fake) picture from myself, ,even my family, but im limited to close-up

MaxiviOP•7/11/24, 3:12 PM

and all your images are cropped at 1024x1024?

Théo•7/11/24, 3:22 PM

ok i understood, to be possible to inpaint in a far away face, you need to Hires a lot, so the Adetailer can actually see the face to inpaint

MaxiviOP•7/11/24, 3:23 PM

use adetailer with correct settings, and with hires fix use 4x ultrasharp or any other good upscalers

Théo•7/11/24, 3:24 PM

my problem was that i wasnt upscalling enough, 1.25 or 1.5 is not good enough for the adetailer to work properly

Théo•7/11/24, 3:25 PM

but 2x upscaling is good, now i have the (body proportion) problem but its easier to fix than not be able to inpaint

MaxiviOP•7/11/24, 4:50 PM

yeah i usually upscale with 1.75x to 2x with 60 steps on adetailer

MMaxivi <@205854764540362752> when you get some time can you try this prompt out on your...

Furkan Gözükara SECourses•7/11/24, 10:51 PM

some prompts will overwrite your likeliness

Furkan Gözükara SECourses•7/11/24, 10:51 PM

like super Mario

Furkan Gözükara SECourses•7/11/24, 10:51 PM

because the model will assosicate it with a certain man

Furkan Gözükara SECourses•7/11/24, 10:51 PM

super Mario - original dataset

Furkan Gözükara SECourses•7/11/24, 10:51 PM

or super man

Furkan Gözükara SECourses•7/11/24, 10:51 PM

it is like ohwx man as brad pitt same

TThéo After 30 training and 5 dataset retake to have the best one, i finally achieve t...

Furkan Gözükara SECourses•7/11/24, 10:52 PM

for mid and distant shots i use adetailer and it works

Furkan Gözükara SECourses•7/11/24, 10:52 PM

you wont get good face with those without adetailer

Furkan Gözükara SECourses•7/11/24, 10:52 PM

because the model inherent training rsolution is too low

FFurkan Gözükara SECourses some prompts will overwrite your likeliness

MaxiviOP•7/12/24, 1:12 AM

i noticed training a little bit of the text encoder 2 helps with that

MaxiviOP•7/12/24, 9:07 AM

@Furkan Gözükara SECourses what is mulit line captioning? the wildcard? what does it do and how do you use it

Khalifa•7/12/24, 12:13 PM

@Furkan Gözükara SECourses do you have any guidelines/guide on how to pick images? as in what type of images should be picked for training?

MMaxivi i noticed training a little bit of the text encoder 2 helps with that

Furkan Gözükara SECourses•7/12/24, 1:41 PM

Nice tips

MMaxivi <@205854764540362752> what is mulit line captioning? the wildcard? what does it ...

Furkan Gözükara SECourses•7/12/24, 1:41 PM

Just write a prompt each line

Furkan Gözükara SECourses•7/12/24, 1:41 PM

It will randomly use during training each one

KKhalifa <@205854764540362752> do you have any guidelines/guide on how to pick images? a...

Furkan Gözükara SECourses•7/12/24, 1:41 PM

I explain in tutorials but I don't have a specific tutorial yet

Furkan Gözükara SECourses•7/12/24, 1:42 PM

I am composing a good dataset to make a dedicated turorial

Théo•7/12/24, 4:43 PM

i tried different dataset style, always perfect quality with different clothing and background

i found out that it is not the best idea to add different facial expression since when using a prompt like "surprised"

if the model use an already surprised expression from the dataset it would be SUPER exaggerated and very weird, even using moderate prompt like "slightly"

I suppose that it would not be an issue if every picture from dataset were captioned, but i havnt found a good tutorial about dataset/captioning and method to use

Also i found that not be able to modify face is a pretty sad, adding a cyborg eye, or facepaint, makup, blood, dirt whatever, tried a lot of prompt even in the adetailer face parameter

I also found out that if your dataset is always with a sharp focused talent and blurry backround, whatever prompt you use, it will always be the same in Stablediffusion, always a sharp focused person, but blury background (i may doing something wrong), i feel like the "from single text file" method is pretty good for something quick, but if you want to achieve God level of training, you need to do the Super anoying long version of data training method

smoke•7/12/24, 8:13 PM

Does anyone have some good results with SDXL trainings and could possibly share results? Mainly for close-ups / portraits.

Tried 1.5 but not too great for me..

daroth•7/12/24, 8:16 PM

what do you mean? theres hundreds of trained good stuff on civitai

smoke•7/12/24, 8:18 PM

Yeah of course but preferably from people here with advice and that can share training parameters

smoke•7/12/24, 8:18 PM

Civitai is full of good stuff ofc

Théo•7/12/24, 10:51 PM

you just follow the onetrainer tutorial from our lord Dr.
1.5 and SDXL works perfect for portrait or close-up, even midrange

The most important is dataset, high quality of 20-25 pictures with everything focused, the most important is eyes/iris

Don't do crazy angle but not only front face picture, add some slightly left/right/pan/tilted angle so you can use full potential of Stablediffusion generation

Use Realvisx4 sdxl model first so you can learn from a model that dsnt need a lot of tweeks to work great, if you want to do NSFW i cannot really help you since i do not produce them, but i assume that using Reg images will break a lot of the capacity of a model to do such generation,

You also gonna need comfy-UI and learn how to use it since you can make different layers of generation to be able to add texture in the face easily (makeup,modification and such) i suppose there is a method to do it without comfy, didn't found it yet.

If you follow Dr. tutorial, copy EVERYTHING and use a GOOD dataset, you will be able to produce even more quality than him (since he is using a medium quality dataset)

MaxiviOP•7/13/24, 1:59 AM

runpod prices decreased a lot for 6000ada and other gpus, looks like they'll try to compete with massed compute

MaxiviOP•7/13/24, 3:22 AM

@Furkan Gözükara SECourses upgraded to new xformers 0.27 and torch, getting way better results and smoother results with xformers and training now on kohya

Sergio A.•7/13/24, 7:56 AM

I'm looking for an AI developer to develop my project.
If u are interested, contact me.

TThéo i tried different dataset style, always perfect quality with different clothing ...

Furkan Gözükara SECourses•7/13/24, 8:23 AM

you got some very accurate findings

MMaxivi runpod prices decreased a lot for 6000ada and other gpus, looks like they'll try...

Furkan Gözükara SECourses•7/13/24, 8:25 AM

true i see A6000 dropped from 0.69 to 0.49 but still we get from 0.31 cent on massed compute with our coupon

MMaxivi <@205854764540362752> upgraded to new xformers 0.27 and torch, getting way bette...

Furkan Gözükara SECourses•7/13/24, 8:25 AM

yes xformers newer versions are way better than older ones

TThéo you just follow the onetrainer tutorial from our lord Dr. 1.5 and SDXL works pe...

smoke•7/13/24, 8:36 AM

Thanks a lot for this! Really interesting I appreciate it.

I’ve done a SDXL training with a pretty good dataset I think, on Realvisxl 4, no captions. But the quality wasn’t that great, and the face resemblance was even worse.

https://discord.com/channels/772774097734074388/1261270101244641302/1261336695110897766

I use Comfy as my main tool but the training itself must be a little bit better I feel like, maybe I’m still doing something wrong

Théo•7/13/24, 8:41 AM

don't forget that you need to use the good vae if not baked-in, Hires 1.5/1.7, adetailer capped on resolution 1024x1024 with a denoising strengh between 0.35 to 0.60,

Théo•7/13/24, 8:42 AM

Also i did an update on my stable diffusion and i couldnt anymore use my training face, , everything worked, but not my trained model, suddently turned bad for no reason, i had to reinstall a fresh sb, dunno why but some ppl had the same pb as me, so maybe you have an adetailer version or something that went wrong

smoke•7/13/24, 9:06 AM

I see! I don't think I even used ADetailer for those but I'll have a look. Thanks a lot!

FFurkan Gözükara SECourses true i see A6000 dropped from 0.69 to 0.49 but still we get from 0.31 cent on m...

MaxiviOP•7/13/24, 9:10 AM

hopefully massed compute gets the 6000ada soon for even faster speeds

MMaxivi hopefully massed compute gets the 6000ada soon for even faster speeds

Furkan Gözükara SECourses•7/13/24, 9:19 PM

i totally agree

Furkan Gözükara SECourses•7/13/24, 9:19 PM

@NicB