it uses same seed by default
it uses same seed by default
Fand yes that is accurate way
Zah ok, what sampling scheduler do you recommend for portraits?
Fstarts with ipn on swarmui
Zso sampler as ipndm, what scheduler?
Zdoing an x/y and want to make sure i'm using the best properties available
Fdefualt
Fdeveloping a new toolkit
Fworking super fast first test 
F
Fi am also removing retina face and replacing it with yolo
Fway faster
Z@Dr. Furkan Gözükara Thanks for all your hard work, following your lesson, i've trained my own model based off of flux dev, and running gguf text encoder and models and made this image
Fi plan to test loras on gguf models
Zhere's another one at 9:16 aspect ratio -
Zi also used your cropper and resizer
Fwow nice
Fi am making gradio for it
Fall in one
GNice art, big pimpin' haha
F--fast brings huge speed boost on rtx 4090
from 1.6 it / second to 2.35 it / second
from 1.6 it / second to 2.35 it / second
Syou could reocmend some model thats work with loras? best regards
DI still don't get the fuss about gguf models? Isn't SwarmUI capable of running the 23GB model on slower GPUs also? Is this to increase the speed of the generations? Or to be able to use more, pin-pointed comfyUI nodes?
Pbut its only on fp8 ? and fp8 degrades LORA likenes
DThat might be true. I'll have to check the settings on massed compute, but perhaps I chose automatic there, and the results were kinda better than what I produce using FP8 on my home PC.
TWhat do you think about training only some layers when doing a Lora? Has anyone tried it?
PHey everyone, quick question, did anyone figure out a way in Flux Dev to get rid of some depth of field blur and out of focus ? I tried various approaches using keywords in the prompt but without any real success. Any advice is much appreciated.
DI hear the feature is not there on Kohya, but possible with AIToolkit, though I am yet to try that.
PBoreal LORA
TYes, is in AIToolKit
PThx, will test that, where can I find that one.. Civit? EDIT: found it i think . Much appreciated
POther quick question, is there a way to test mutliple loras in comparison in Stable Diffusion Forge webUi via script , for example X/Y/Z matrix or similar?
TYes you can do with xyz and select prompt S/R
Pin forge it works same as in A1111 and in comfy its also posible but its hard to make it work
PIs there anything to watch out for if you say its hard to make it work ?
PWhat are the best settings on a rtx 3090? feels some times very slow and other times it runs smooth ... bit weird
Pim just not good at comfy. it could be easy. but i couldnt figure it out...
PSame here.. either render gets stuck or doesnt work.. in forge.. not good in comfy better said comfortable to try that, and setup a workflow.. just want to add more realism and compare the loras for best result.. maybe i check swarm UI ?
or doesnt work.. in forge.. not good in comfy better said comfortable to try that, and setup a workflow.. just want to add more realism and compare the loras for best result.. maybe i check swarm UI ?
I
Swe add "<segment:face,0.7>photo of" in the very end of our prompt?
Sto get a better quality face?
DYes, the face gets detected, segmented, and recreated, 'cuz you may not get high quality faces in distant (half/ full body) generations
DI find the results for this better than the yolo one.
Fye i plan that
Ftrue rtx 4090 cant run fp16 - not enough vram
Fit is. but gguf is i think more for lower vram much lower
Fi am waiting kohya for this
Fyes it inpaints face
Flike adetailer
Fyou have install error