Software Engineering Courses (SECourses)•11mo ago

Training for 5 minutes?

druhl•2/6/25, 4:04 PM

@Dr. Furkan Gözükara I had tried 'couple trainings' with images of them together in the input dataset. I captioned images using joycaption. Tbh, the model always learnt one person better than the other. What could be the reason for this? I used the same settings as your one -character dreambooths.

RRenzo Training for 5 minutes?

Furkan Gözükara SECourses•2/6/25, 4:07 PM

yes

Ddruhl @Dr. Furkan Gözükara I had tried 'couple trainings' with images of them together...

Furkan Gözükara SECourses•2/6/25, 4:07 PM

interesting not sure why

FFurkan Gözükara SECourses yes

RenzoOP•2/6/25, 4:08 PM

Targetting a few layers?

RenzoOP•2/6/25, 4:08 PM

What's the size of the LoRA file?

RRenzo Targetting a few layers?

Furkan Gözükara SECourses•2/6/25, 4:08 PM

nope fully layers

Furkan Gözükara SECourses•2/6/25, 4:08 PM

128 rank

Furkan Gözükara SECourses•2/6/25, 4:08 PM

fp16 saving : 1.27 GB

RenzoOP•2/6/25, 4:09 PM

That's a small file.

RenzoOP•2/6/25, 4:09 PM

Resemblance is good?

RenzoOP•2/6/25, 4:10 PM

Oh, and how fast is the H200?

RRenzo Oh, and how fast is the H200?

Furkan Gözükara SECourses•2/6/25, 4:51 PM

really fast

RRenzo Resemblance is good?

Furkan Gözükara SECourses•2/6/25, 4:51 PM

not too bad but lower quality here

Furkan Gözükara SECourses•2/6/25, 4:52 PM

Furkan Gözükara SECourses•2/6/25, 4:52 PM

left ours right faster

Furkan Gözükara SECourses•2/6/25, 4:52 PM

you see faster is overfit to get decent resemblance

Furkan Gözükara SECourses•2/6/25, 4:53 PM

Furkan Gözükara SECourses•2/6/25, 4:53 PM

Furkan Gözükara SECourses•2/6/25, 4:54 PM

i will try 2x gpu next with 1024px

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Bro what you are doing is impressive I didn't quite understand your pixel chunks...

.tarkan•2/6/25, 5:16 PM

You can try out the model here. https://civitai.com/images/56138354

..tarkan You can try out the model here. https://civitai.com/images/56138354

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/6/25, 5:18 PM

Thanks but I want to know exactly how are you doing this

FFurkan Gözükara SECourses not too bad but lower quality here

RenzoOP•2/6/25, 6:19 PM

Yeah. The original one is excellent. How many images did you use?

RRenzo Yeah. The original one is excellent. How many images did you use?

Furkan Gözükara SECourses•2/6/25, 9:10 PM

dataset shared here : https://www.patreon.com/posts/114972274

Patreon

Perfect Quality Example Model Training Images Dataset - Can Be Used...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

Furkan Gözükara SECourses•2/6/25, 9:10 PM

28 images

Furkan Gözükara SECourses•2/6/25, 9:10 PM

it is 170 th epoch

Maki99•2/6/25, 10:14 PM

hehe

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Thanks but I want to know exactly how are you doing this

.tarkan•2/6/25, 10:44 PM

I've written python scripts with chatGPT that will process the high resolution images for training by cutting them into 1024x1024 pieces with a bit of overlap.

..tarkan You can try out the model here. https://civitai.com/images/56138354

TheToday•2/6/25, 10:46 PM

Amazing, can you upload the model to tensor art please?

.tarkan•2/6/25, 10:46 PM

Sure!

..tarkan Sure!

TheToday•2/6/25, 10:46 PM

Thank you!

.tarkan•2/7/25, 1:10 AM

@TheToday Done! https://tensor.art/models/827274559054055826

Flux Sigma Vision Alpha 1 - 1 | Stable Diffusion Model - Checkpoint

0 runs, 0 stars, 0 downloads. This fine tuned checkpoint is based on Flux dev de-distilled thus requires a special comfyUI workflow and won't work very well ...

.tarkan•2/7/25, 1:10 AM

Little comparison to see the improvement side by side.

MMaki99 hehe

AIGambino•2/7/25, 1:45 AM

Is there a resolution limitation of this?

MMaki99 hehe

SirPoopsAlot•2/7/25, 2:32 AM

you recreated this? holy moly! do you have a breakdown on how you did this?

AAIGambino Is there a resolution limitation of this?

Maki99•2/7/25, 6:59 AM

not really the higher the resolution the longer it takes to render obviously - and you get better results but also diminishing returns eventually - this was resized to 1080p i believe took 5min to render

SSirPoopsAlot you recreated this? holy moly! do you have a breakdown on how you did this?

Maki99•2/7/25, 7:02 AM

i sent the workflow above - input video - play with the prompt and setting, make sure to match the skip steps with the normal scheduler step to equal 20 step of genning anything under that breaks teh genning

Maki99•2/7/25, 7:02 AM

and prompt basically everything betwheen the both videos i wish theres an llm that can do that for me to be more accurate

wardensc2•2/7/25, 7:25 AM

I already test some top Flux dev de-distilled model, quality quite dissapointed, most of them only good on realistic image, for the style only the same or even worse than Flux dev although time for render is double due to CFG >1. Are there any de-distilled models you guys can recommend and maybe can be used to train lora and finetuning ?

MMaki99 i sent the workflow above - input video - play with the prompt and setting, make...

SirPoopsAlot•2/7/25, 7:37 AM

Awesome, thanks a ton Maki

FFurkan Gözükara SECourses deep face labs model?

masood•2/7/25, 12:08 PM

yep

masood•2/7/25, 12:08 PM

and whats the best tool for lip sync?

Hael•2/7/25, 3:36 PM

What would you guys recommend for a flux checkpoint training on Kohya for a character model.. i have 40 images on a 3090ti (24 gb, 64 gb of ram). More than batch size 1? Is current best practice not to caption our training sets and to use ohxw woman? Things change so fast in this world?

AiInfluence•2/7/25, 6:01 PM

@Dr. Furkan Gözükara I'm currently using Flux Q8 on Kaggle , is there something better I can use for multi purpose? (realism,style etc). I saw in chat something about de distilled versions?

Mmasood and whats the best tool for lip sync?

Furkan Gözükara SECourses•2/7/25, 9:03 PM

nothing i know specific

HHael What would you guys recommend for a flux checkpoint training on Kohya for a char...

Furkan Gözükara SECourses•2/7/25, 9:03 PM

batch size 1 is best

Furkan Gözükara SECourses•2/7/25, 9:03 PM

we have 24 gb config do fine tuning / Dreambooth