i use a40 what rank should i use ?
i use a40 what rank should i use ?
Fthen very likely
Frank 1 best
Fbut it is too slow if you need speed
Frank 2
Fi feel like with consistent dataset no captions worked better than captions
Ffor style for flux training
Flet me share link with you
Li did not use any caption files and also no reference images -- just 100 epochs with repeat set to 1 having 132 images to train
Lhowever i notice the vertical lines are "lesser" with lower initial resolution like 1024x512 (instead of something like 1400x700)
Lwhat generation resolution do you recommend in common?
Fmake all train iamges 1024x1024 and generate 1024x1024
Fthen you can try multi resolution
Fcurrently people reporting multi resolution with kohya not working very good for flux - i am not sure - but as i said first go with 1024x1024 and once you have some decent results you can try multi res training and compare
Lor maybe it is because of my pics are being oversized for training? Or is it rather the mixed aspect ratio?
Fmixed aspect ratio
Fyou have too many
Fand too little number of images
LI have 132 images for lora training -- I thought that even would be too much -- as other people I heard of only train with like 20 images? So 132 is commonly ok?
PI used rank 2 and it worked. But the problem is that I used 25 images but I don't know why the num train images * repeats is up to 1025. Am I setting it wrong somewhere? And if so, where should I go to fix it?
Fmore better
Fmake repeat 1
Fif you have another folder in training folder
Fit uses all
Pthanks you
DI just made the payment, and I hope you can help me achieve the goal of generating comic images of my NFTs. I’m starting to watch your video. Best regards.
F
Fstyle training concluded i written everything there
LSo to ensure -- it would be ok to have mixed resolution for training images but they should have the same aspect ratio
For example havin mixed 1000x1000, 1200x1200, 1400x1400 and so on but all have the aspect ratio of 1:1
Can I have the some with aspect ratio 2:1 like havin 70 images with 2000x1000 and 70 with 1000x1000?
For example havin mixed 1000x1000, 1200x1200, 1400x1400 and so on but all have the aspect ratio of 1:1
Can I have the some with aspect ratio 2:1 like havin 70 images with 2000x1000 and 70 with 1000x1000?
LBackground: I do not want always only to crop the face -- because the lora should also "learn" the body proportions
Fthis sounds better than your older case
Li´ll try, thank you -- by the way - I also subscribed to your patreon cause I think you do a great work, thank you!
FFull Training Tutorial and Guide and Research For a FLUX Style Hugging Face repo with all checkpoints, prompts and more details > https://huggin...
D@Dr. Furkan Gözükara I’ve read in detail what you did with MonsterMMORPG/3D-Cartoon-Style-FLUX, and now I’ll try to train following that guide. I have 800 images of 2608 x 2608; they are all spheres that vary in accessories and country color. My goal, after achieving a decent result that reuses the character in different scenarios, is to deploy it on an API that I will later query through a Telegram bot. One question I have is whether to use Massed Compute or RunPod?
Fgo massed compute
Fway more dependable
Frent 4x A6000 - 1.25 usd per hour
Fprobably you can get good results with early epochs if they are consistent
F66 images required only 75 epochs
DI have 800 different characters, between accessories and country, do you think that training with the 800 images is not the way to go?
I should choose 100 and work with them?
Because per country I have 4-6 versions which share the same flag color, but different accessories.
I should choose 100 and work with them?
Because per country I have 4-6 versions which share the same flag color, but different accessories.
DThe least I hope in this experiment is to get to occupy a prompt, that I can pass it for example “Ball on the beach” I hope one of those balls with a background on the beach.
Fas long as they consistent and not repeating i would say use all
Falso in your case captions may work better not sure though
XThis is interesting is it an nft project?
Dyes.
DDo you recommend using them without background?
Fi say keep dont remove
Fit is how you want to generate right?