is it possible to train with just one image and have consistency in character?
is it possible to train with just one image and have consistency in character?
F2 step. so train generate and train again
Fbecause single image usage not enough
XI’m geting an error while downloading Training_model_files.The error message says: 'Could not set the permissions on the file... Error: [Errno 13] Permission denied.' The download continues, but I’m not sure if this will cause issues later.
J@Dr. Furkan Gözükara or anyone else have tried fine-tuning flux on a smartphone photo style? So that when afterwards doing a new fine-tuning on a person, the photos will look more casual/realistic?
I'm trying right now with 170 mixed photos, most from different Instagram profiles, maybe 2/3 of them includes people (not sure if that's smart or not) and otherwise just using Furkans default 48GB Batch size 7 config.
No idea how many steps/epoch is needed, or what the optimal learning rate is.
Using the word "style" as the Class Prompt, also not sure if that's smart.
I'm trying right now with 170 mixed photos, most from different Instagram profiles, maybe 2/3 of them includes people (not sure if that's smart or not) and otherwise just using Furkans default 48GB Batch size 7 config.
No idea how many steps/epoch is needed, or what the optimal learning rate is.
Using the word "style" as the Class Prompt, also not sure if that's smart.
Frelated to your windows folder permissions
Fyou can use as smartphone photo
Fbut more images may weork better
Xwill it affect me during traning?
Fit shouldnt but who knows
Fif windows prevents python to write
For antivirus
Fbut i think should work
F@The90sPope
F
Fthis is how you use FLUX dev on Forge. I generated this image on RTX 5090
F
TI have stripped down system with no antivirus, it's an overclock build for breaking records haha, no antivirus. barebones windows, lets see...
Fbest case
Ti have the scaled model it seems, im on 16gb vram
Ti have fp8
Tif i copy the settings, and try different sampler scheduler combos that are known to work, still black and dots. Tried them all basically. No errors, nothing.
Thmm
Tbecause i actually get an image, it's just not what I prompted, its black image with green dots, no matter what i type in prompt
Feasy your model not supported by forge
Taah, okey, thanks, gonna try something else, what does text inside image good?
Tideogram is what i use now, but it's paid
Fflux is next best
Fflux 1 dev
Toption 10 on download models script
Twill try that
F2 should be better
Fdid you try?
Fif 2 failed
Fyes then 10
Tyea, 2 failed
Tthat's what i have been trying now
Tdownloading 10 and give it a spin
Tthanks for help, i take it from here 
Fyou are welcome. forge is not getting updated that is biggest issue 
Fi am downloading scaled to test myself
Foption 1 should also work fine with forge optimizations i presume
Tnice, btw your wan 2.1 1.3B gives me better results than kijai nodes on comfy
Fnice. yes kijai does so much optimizations and trade of is quality
Fi am waiting official teacache because of this reason also
TI just but flow to 8 instead of 6 since devs say 8 to 12 is best on 1.3b
Fi can change default
Fwhich one is best you think?
F8?
TI use 8 all the time, but it really depends, lottery