Yes i already done it. When extract the highest extractable rank which is 640 the size of lora is si
Yes i already done it. When extract the highest extractable rank which is 640 the size of lora is similar to fp8 checkpoint or gguf q8 full model
Acan anyone share best practices for training and generation to achieve maximal likeness of real person and generated images?
U
Fthis video is the best one you can find https://youtu.be/FvpWy1x5etM
YouTubeSECourses
If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...
Fvideo is really high res
UThank you, Dr. Furkan!
Xcurrently trying to train a character lora (Anime, Pony checkpoint), but it also get that animescreen-cap style(usual anime style), wonder if there is a way to train character without those styles and effects
DHi @Furkan Gözükara SECourses, I saw you commented on Boris Noll's LinkedIn post. I'm curious if you have explored these methods of erasing concepts from the Flux base model, and training it for higher resolution output? I would love to learn more about this workflow (possible also his object and concept LoRA training methods). Thank you.
DWhat's the resource you're referring to for style training the 3D-Cartoon-Style-FLUX on Hugging Face?
Fi think that post is not totally honest :d
DThanks for your reply. Hmm really, the base model part or regarding the loRA's? Would you mind to ellaborate?
Fthe erasing concepts doesnt make sense
Fand i dont know how he made such script himself
Fit is so hard to make such scripts accurately
DOkay I see. Curious if he'll ever share that script then, What about finetuning the base model to output a higher resolution (2560x2560px)?
Fwell i had tested on myself limitedly and i didnt notice improvements
Fbut when training a car i dont know
Fit also sounds unreal it is over 5 mega pixel
Fmodel is 1 mega pixel
DThanks for your insight. I was also very suprised about this post since I have never seen something like that before.
Regarding the concept LoRA's I assume he trained 3 style loras for additional "aumotive" looks?
Regarding the concept LoRA's I assume he trained 3 style loras for additional "aumotive" looks?
Fi am not sure
DOk thank you. Will you dive into more product/object specific (lora) training experiment/tutorials in the future you recon?
Fi may make but too many jobs waiting i need to research sd 3.5 
Fcurrently testing model varients impact on lora
Fand lora extraction
Fto post on civitiai
DHaha makes sense! We'll keep following your work closely, thanks!
T@Furkan Gözükara SECourses Is there any reason to use kohya over OneTrainer quality-wise for Flux? Is it easy to convert your best koyha config settings to OneTrainer? Also noticed onetrainer supports SD3.5 There are some nice Quality of life things in Onetrainer, like dynamic save anything ondemand, resume from backup, changing samples during training, sample on demand etc. thoughts?
WKohya have most of your thing but you need to learn how to do it. I use Kohya since SD 1.5 never feel that I need to use OneTrainer. Koyha maybe a bit more technical but when you understand it, it's very versatile
TI like Kohya, it just needs some of the quality of life things. like save now.. easy resume, change sample prompts. So many times i want to use GPU for something, but have to wait 20 epochs 
would be nice to have a bit more control..
would be nice to have a bit more control..FKohya also supports 3.5 . also you can convert my configs
FOneTrainer was lacking previously
Fhopefully I will compare and publish configs
Wbut I need to study again for someone who already know SD for a year or more, people tend to become more lazy especially when Onetrainer does not have something extraordinary than Kohya and also very few people like @Furkan Gözükara SECourses who dedicate time to study Kohya. Finding the right config for lora training or fine tuning is really time consumption and require deep knowledge about AI programming
Fyou can clone our kohya configs to onetrainer i did it for sdxl and sd 1.5 :d
Wreally, how about training speed Kohya vs Onetrainer
Fi havent compared yet but kohya also just made a branch to speed up fine tuning
ZCould you get in touch with me so we can discuss this?
TIs this included in your new v11 zip what is the difference? Do I need to upgrade?
Fnot included yet
Fi need to test first
Fbut upgrade it helps me :d
Walready messsage for you
Fis it worth training on higher resolutions then 1024 x 1024 ? lets say like 3k? 4? comething like 3000 x 2000 ? Thanks
Flast time i tested i didnt see benefit
FHello again,batch 7 at 250 images on my 4090 is suuuuuper slow
F
Fwas able to spin a 250 last night
Fwhat am I missing?
Fmessage.txt14.16KB