Tried it fal.ai very low quality lool
Tried it fal.ai very low quality lool
BReally? You mean it produced very low quality results?
YYes it did !
YOver grained face
Y+
YSimilar to fast training the portrait output face isn't
Yresembling well
BSure it wasn't your data set?
YSame dataset using kohya produces better results
Ehmm I cant see much difference--
Ety for this effort tho
Fyep
Fbut 30 better
Flol and expensive too
Fpatreon ranks updated by cumulative membership - you have to have connected your discord on patreon and be active
FFLUX image generations comparison speed
F
�Is facefusion 3 the absolute best faceswapper? I'm helping a girl who has been using a bad faceawapper, resulting in her training pics lacking facial details. So I was thinking of first swap all the faces again, with her base face, which is of high quality.
Fnone of these will ever yield training results
Fbut facefusion is best 0-shot atm
Y
YWhat do you think of this output !
YNo supir nothing pure lora
YKohya fine tunning
Y100 times better than fal
DYes, it's a great image. 
DThat's spooky 
Elol
�yes of course, Dr. Thank you. I wanted to improve the training pictures before starting training.
SImpressive
TFlux?
A@Dr. Furkan Gözükara urgent request to make an F5 TTS Fine-tuning tutorial + massed compute installer I cannot stress enough how good it is !
AWe need something valuable for voice too
MToday I tried training flux lora with 30 images using A6000 and at 9 s/it it's giving me 11 and a half hours. With 4090 at 5.11 s/it (and the same number of images) it gave me 6 and a half hours lol. Why is that? Oh and I used the 24GB_GPU_Quality_Tier_2_23100MB_9.8_Second_IT_and_T5_Attention config in both cases. Is it because of "--blocks_to_swap 12"?
�if all training pics are 2048x2048, should I keep them that way, or reduce to 1024x1024?
�Use tier 1 when using a6000 for max quality
MAs far as I know Tier 1 only adds Train T5-XXL
Myeah it says it in the Configs Explanation Must Read File. Although I probably should have selected the 48gb config file not the 24gb
YYES
Y
Y@TheToday all these are flux lora
Y
YGuys am launching a backend kohya trainer + inference that outputs these images ! Anyone would like to test it for free ?
Y
YI now face an issue with jewelery and accesories they are not high quality do you guys know any way to optimize them ?
Y
Y
Tyes I want
Y
DTbh, jewelery works best with direction. Like img to img. Without it, the model is just lost and breaks things here and there.
DMaybe use a jewellery lora