Software Engineering Courses (SECourses)•15mo ago

simpler grids without reg images

Ddxqb for those who were interested in this https://discord.com/channels/7727740977340...

J2_Nedo•10/11/24, 5:51 AM

some vips are already trained in to some degree, in my experciene no reg, no prompt gives me more then good enough results. and i switched to the competitor product (fgym) because it is so easy to use

JJ2_Nedo some vips are already trained in to some degree, in my experciene no reg, no pro...

dxqbOP•10/11/24, 5:56 AM

"no reg, no prompt" will work but change the entire model. Barack Obama will look like your trained subject.

Ddxqb "no reg, no prompt" will work but change the entire model. Barack Obama will loo...

J2_Nedo•10/11/24, 6:14 AM

i dont train known people, just ohwx works super

JJ2_Nedo i dont train known people, just ohwx works super

dxqbOP•10/11/24, 6:29 AM

you don't get the point, please read the thread that I linked above first.

Ddxqb you don't get the point, please read the thread that I linked above first.

J2_Nedo•10/11/24, 6:52 AM

its about the bleeding i know, if it works it could be usefull, but its only for onetrainer right now?

dxqbOP•10/11/24, 6:54 AM

yes. SimpleTuner wants to implement, and I've heard kohya too, but nothing ready

Ddxqb simpler grids without reg images

Furkan Gözükara SECourses•10/11/24, 1:21 PM

great results but i couldnt get it :d

theJindesan•10/11/24, 1:39 PM

I have a few questions. I'm now trying the DB method for flux training, no captions, 30 images, 1 repeat, 150-200 epoch about 5000-6000steps - is that too much. Also since i'm a male but i have long hair, every sample image during training shows as a women, in my case is it better to provide captions saying I'm a man? Joy caption also thinks I'm female. Also i noticed that in any of my trained/extracted lora, i can never add any other person in the picture, because they all look like me, even though I'm using a numeric trigger, and even if I import another celebrity lora, it just blends the 2 lora's together. is there no way to do an AI image of myself and my favorite celeb? ie. "photo of k3v1n with billie elish" (another lora)

TtheJindesan I have a few questions. I'm now trying the DB method for flux training, no capt...

Furkan Gözükara SECourses•10/11/24, 1:43 PM

about second part yes it bleeds

Furkan Gözükara SECourses•10/11/24, 1:43 PM

@dxqb found a way to lessen it on onetrainer

Furkan Gözükara SECourses•10/11/24, 1:43 PM

but i couldnt make it

Furkan Gözükara SECourses•10/11/24, 1:43 PM

about first part only use

Furkan Gözükara SECourses•10/11/24, 1:43 PM

ohwx man

Furkan Gözükara SECourses•10/11/24, 1:43 PM

it should work with sufficient training

Furkan Gözükara SECourses•10/11/24, 1:43 PM

i will update patreon post with grids and my best epochs

Furkan Gözükara SECourses•10/11/24, 1:43 PM

based on number of images

Furkan Gözükara SECourses•10/11/24, 1:43 PM

and batch size

FFurkan Gözükara SECourses about second part yes it bleeds

theJindesan•10/11/24, 1:46 PM

so inpainting is the only way? how does it work in base flux compared to lora? because in base flux you can specify 2 embedded base celebrities, ie. john wick and taylor swift playing in a concert together.

TtheJindesan so inpainting is the only way? how does it work in base flux compared to lora? ...

Furkan Gözükara SECourses•10/11/24, 1:46 PM

the problem is that since we train on distilled model

Furkan Gözükara SECourses•10/11/24, 1:46 PM

it bleeds a lot

Furkan Gözükara SECourses•10/11/24, 1:47 PM

but famous people were trained before the distillation

Furkan Gözükara SECourses•10/11/24, 1:47 PM

so flux de-disttiled models may likely to fix this problem

Furkan Gözükara SECourses•10/11/24, 1:47 PM

when they are ready

FFurkan Gözükara SECourses when they are ready

theJindesan•10/11/24, 1:49 PM

thank you i understand. So for now, there is no way for me to train me and my friends and create pictures together, even with unique words in a checkpoint? could i train them all at once into a new checkpoint or would it bleed? ie. 1_k3v man, 1_fr1endone, 1_fr1endtwo (ie each friend is a concept)?

TtheJindesan thank you i understand. So for now, there is no way for me to train me and my f...

Furkan Gözükara SECourses•10/11/24, 1:50 PM

there is a way

Furkan Gözükara SECourses•10/11/24, 1:50 PM

have yourself in same picture

Furkan Gözükara SECourses•10/11/24, 1:50 PM

and train on that :d

Furkan Gözükara SECourses•10/11/24, 1:50 PM

then it will work

theJindesan•10/11/24, 1:54 PM

Thank you for all of your help and community contribution, I'm at a university in Canada, and we're looking at upgrading our old Tesla T4 servers in our engineering dept, so we can do more AI research. we're a bit behind.

TtheJindesan Thank you for all of your help and community contribution, I'm at a university i...

Furkan Gözükara SECourses•10/11/24, 1:55 PM

awesome

dxqbOP•10/11/24, 4:08 PM

@Furkan Gözükara SECourses what were your results with trying to train Flux on your reg data set? I'm attempting something similar

dxqbOP•10/11/24, 7:12 PM

@Furkan Gözükara SECourses please see here https://discord.com/channels/1102003518203756564/1149379539454140538/1294374243060027475 . I am not entirely sure this has been the reason your config didn't work, but it was the reason someone else's didn't work.

Ddxqb <@205854764540362752> what were your results with trying to train Flux on your r...

Furkan Gözükara SECourses•10/11/24, 8:23 PM

never worked

Linda Watson•10/12/24, 8:37 AM

Please @Furkan Gözükara SECourses respond to DM

Ddxqb simpler grids without reg images

dxqbOP•10/12/24, 9:46 AM

this is implemented in SimpleTuner now

LLinda Watson Please <@205854764540362752> respond to DM

Furkan Gözükara SECourses•10/12/24, 10:27 AM

checking

Ddxqb this is implemented in SimpleTuner now

Furkan Gözükara SECourses•10/12/24, 10:28 AM

have you tested there?_

FFurkan Gözükara SECourses have you tested there?_

dxqbOP•10/12/24, 10:28 AM

No, never used SimpleTuner so it would be a lot of effort for me to try

dxqbOP•10/12/24, 10:29 AM

but checked the code. they do exactly the same thing

dxqbOP•10/12/24, 10:35 AM

other people are

Ddxqb other people are

Furkan Gözükara SECourses•10/12/24, 10:48 AM

i will wait kohya

daroth•10/13/24, 5:37 PM

The image is a high-resolution photograph featuring a woman with a stern expression, likely in her mid-30s, standing in a dimly lit, industrial setting. She has shoulder-length, straight blonde hair with a slight wave, and her skin has a light, natural complexion with a touch of blush on her cheeks. Her eyebrows are arched, and her lips are slightly pursed, conveying a look of determination or frustration. She is wearing a bright blue zip-up jacket, which is unzipped to reveal a hint of a purple top underneath. The jacket appears to be made of a smooth, synthetic material with a slightly shiny texture.

Civitai auto captioner - its free

daroth•10/13/24, 5:38 PM

Lots of filler but otherwise reads solid enough

theJindesan•10/15/24, 1:57 AM

@Furkan Gözükara SECourses have you tested the 6GB and 8GB configs for DreamBooth Flux training? For some reason i can't get it working on my 2080 super 8GB ram. I've tried both 6 and 8GB. Both crash as it starts the first step. Either out of memory or another error. I have no trouble on 3090. Also on the same PC with 2080 super, i can run fluxgym with same settings and it works.. But its not using DB method. "RuntimeError: CUDA error: out of memory
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with