Software Engineering Courses (SECourses)•17mo ago

It does work, just need the right batch size

GGomezBro whats that

Furkan Gözükara SECourses•9/1/24, 10:37 AM

I give 1 to 1 lecture

TThereiver It does work, just need the right batch size

Furkan Gözükara SECourses•9/1/24, 10:37 AM

Batch size 1 best

ThereiverOP•9/1/24, 10:38 AM

For quality I agree

Furkan Gözükara SECourses•9/1/24, 10:38 AM

Batch size 2 made like %25 improvement

Furkan Gözükara SECourses•9/1/24, 10:38 AM

Rest didn't do

Furkan Gözükara SECourses•9/1/24, 10:38 AM

I tasted on a6000

GomezBro•9/1/24, 10:40 AM

even batch size 4 works on A6000 bros I tested it today

ThereiverOP•9/1/24, 10:40 AM

Thanks - will keep that in mind! Yeah it would be awesome if you could make a config for prodigy, I was able to train the darkest dungeon sdxl Lora in probably 1/10th of the time

GomezBro•9/1/24, 10:40 AM

RANK 5

TThereiver Thanks - will keep that in mind! Yeah it would be awesome if you could make a co...

ThereiverOP•9/1/24, 10:40 AM

I will have a look to see if I can find my old config

GomezBro•9/1/24, 10:40 AM

I got it to over 30+ iterations

GomezBro•9/1/24, 10:41 AM

it has a lot of VRAM i figured what the hell and tried it

ThereiverOP•9/1/24, 10:43 AM

Nice one man

ThereiverOP•9/1/24, 10:51 AM

Found my old prodigy config - will do some testing with flux dev and let you know if it's any good. Note I'm training a style, so the dims will be higher

TThereiver Found my old prodigy config - will do some testing with flux dev and let you kno...

GomezBro•9/1/24, 10:58 AM

brother i want to do that too, how many images will u train?

GomezBro•9/1/24, 10:58 AM

please do share infos

ThereiverOP•9/1/24, 10:58 AM

350 images in dataset roughly

GomezBro•9/1/24, 10:58 AM

i want to train an NFT collection into a style

ThereiverOP•9/1/24, 10:59 AM

This was my lora https://civitai.com/models/65324/darkestdungeon

DarkestDungeon - v2.1 | Stable Diffusion LoRA | Civitai

v2.1 Updated 19/5/23 If you enjoy the models I create, please consider buying me a coffee! Ko-fi I am pleased to announce the release of DarkestDun...

GomezBro•9/1/24, 10:59 AM

why so little if you dont mind me asking bro?

ThereiverOP•9/1/24, 10:59 AM

350 is quite a lot for a lora tbf

TThereiver This was my lora https://civitai.com/models/65324/darkestdungeon

GomezBro•9/1/24, 10:59 AM

thats a while back cool

GomezBro•9/1/24, 10:59 AM

Dr just told me to use more

GomezBro•9/1/24, 10:59 AM

to do 10K over 1K images

ThereiverOP•9/1/24, 11:00 AM

I would say 10k for a model not a lora but thats my opinion. 2000 for a style lora is a good amount. but you can get away with less

ThereiverOP•9/1/24, 11:01 AM

also mine: https://civitai.com/models/124498/darkest-dungeon-sdxl-lora

Darkest Dungeon SDXL Lora - v1 - Main Release | Stable Diffusion Lo...

I am pleased to announce the release of DarkestDungeon SDXL Lora What is it? Darkest Dungeon SDXL is a lora intended to capture the artstyle from '...

ThereiverOP•9/1/24, 11:01 AM

It's only right I make one for flux

TThereiver I would say 10k for a model not a lora but thats my opinion. 2000 for a style lo...

GomezBro•9/1/24, 11:07 AM

the thing is i want to make the quality of that NFT very closely, so close that it looks like it with the prompt i create

GomezBro•9/1/24, 11:07 AM

so i want it to be very accurate and high quality

ThereiverOP•9/1/24, 11:08 AM

Yeah it's doable. You just need a consistent dataset

ThereiverOP•9/1/24, 11:09 AM

Mine for instance has about 8 different concepts within

ThereiverOP•9/1/24, 11:09 AM

Whereas an NFT would usually only have one or two

ThereiverOP•9/1/24, 11:21 AM

I was working with a guy that pretty much mastered the '100 steps or less' challenge, I had a config somewhere for that too. Will see if I can find it. Might have been a TI though

Furkan Gözükara SECourses•9/1/24, 11:24 AM

i will be away for a while can check messages later

ThereiverOP•9/1/24, 11:26 AM

All good man

FFurkan Gözükara SECourses prev vs new

TOPUI•9/1/24, 12:16 PM

When will this latest JSON configuration file be updated?

ThereiverOP•9/1/24, 1:01 PM

Can confirm batch size 1 is fastest (in my case anyways)

sergio_gonzo•9/1/24, 1:59 PM

Dr Furkan the loras FLUX could work in any checkpoint or must be in specific ones?

Afroman4peace•9/1/24, 2:12 PM

hi all

Ssergio_gonzo Dr Furkan the loras FLUX could work in any checkpoint or must be in specific one...

ThereiverOP•9/1/24, 2:14 PM

I believe flux dev loras only work with flux dev, and flux schnell loras for flux schnell. AFAIK they're not compatible with SD 1.5, SDXL etc

sergio_gonzo•9/1/24, 2:24 PM

thank you man, but i saw there is a variant in size of flux dev i must see sometinh like that, if have f16 maybe?

Ssergio_gonzo Dr Furkan the loras FLUX could work in any checkpoint or must be in specific one...

dxqb•9/1/24, 2:29 PM

I tried a character lora on the "New Reality" FLUX preAlpha checkpoint. results weren't good

dxqb•9/1/24, 2:30 PM

so probably same as SDXL. some work better than others, but generally Loras don't translate well between checkpoints

Afroman4peace•9/1/24, 2:35 PM

I am a checkpoint trainer myself (Colossus Project XL) yes some checkpoints are trained in very unique ways that may can result in bad trainable models

AAfroman4peace I am a checkpoint trainer myself (Colossus Project XL) yes some checkpoints are ...

dxqb•9/1/24, 2:37 PM

can I ask you an unrelated questions?
what do you guys do that somehow screws up the possibility to train embeddings on your models? I haven't tested Colossus, but on all other finetuned SDXL checkpoints like RealVis, Juggernaut, ... you cannot train embeddings, only on SDXL base. Same with Pony: you can train embeddings on PonyDiffusionXL, but not on any derivates like PonyRealism

Afroman4peace•9/1/24, 2:41 PM

most of all many guys in the community loves to copy stuff. At the beginning of SDXL I was the first that combined the refiner model with the normal model. I think the first one was my Hephaistos model... not long after that nearly everyone used it.

Afroman4peace•9/1/24, 2:42 PM

Then there are some sdxl models that have seperated from the original so much that its nearly something completely different than sdxl base

dxqb•9/1/24, 2:43 PM

I don't mean that it's different or that the embedding doesn't translate well between checkpoints. it's impossible to create an embedding on finetuned SDXL models

Afroman4peace•9/1/24, 2:46 PM

I haven't trained an embedding on my models. There was simply no need to do it. I may will test this

Afroman4peace•9/1/24, 2:51 PM

@Dr. Furkan Gözükara I am here to say thanks for your tutorial and the kohya_ss settings. I will use this to train new Flux models. I have just a 4070ti with 12GB so this will be a great help for me

It does work, just need the right batch size

Similar Threads

Similar Threads

Similar Threads