--fast brings huge speed boost on rtx 4090 from 1.6 it / second to 2.35 it / second
--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
Fyou probably didnt follow this video
FYouTubeSECourses
Welcome to the Pillars of AI Application Libraries: How to Install Tutorial! 
Are you ready to embark on a journey of installing AI libraries and applications with ease? 
In this video, we'll guide you through the process of installing Python, Git, Visual Studio C++ Compile tools, and FFmpeg on your Windows 10 machine. 
We'll also show you ho...
Are you ready to embark on a journey of installing AI libraries and applications with ease? In this video, we'll guide you through the process of installing Python, Git, Visual Studio C++ Compile tools, and FFmpeg on your Windows 10 machine. We'll also show you ho...F10 epoch
Fbut that lora training brought realims is not there yet
PDinos still need to get better .. .. awesome image
.. awesome imageF
Fwithout prompting it generates smiling
F
F20 epoch
Fbut i see the impact of poor lightning / focus of dataset
F
F
Fdmn model learnt even my broken teeth haah
Fi will now retrain batch size 1 with up to 30 epochs
Fafter 30 epochs it became cooked
Fmaybe maximum 40 epochs
Fit learnt my body shape too perfectly
F
FT5 attention mask + T5 XXL speed is 4.52 second / it on L40S - training everything
Dfor generation is Euler Simple still the best bet?
FiPNDM is best i think
