@Dr. Furkan Gözükara what is the best speed I can get full finetuning with my 3090ti?
@Dr. Furkan Gözükara what is the best speed I can get full finetuning with my 3090ti?
EThanks this is really encouraging to see. People have been telling me you only need 2000 steps or it is overtrained, but most of my attempts look like your examples. 2000 steps PS3 style and closer to 6000+ start to look more photographic. BTW could you make a guide on how to tell if you have over or undertrained if you haven't already?
F10 second / it
Fit is decent believe me worths it
Fyes i will make a guide
Factually preparing grids right now
Fwill share all grids on patreon post
Fwith explanation
OCurious what your conclusions will be on batch sizes, you seemed to be getting good results with 7
HHow long would it take to train a Flux LoRA with 105 images? I'm using an RTX 3090
Hwith Dr. Furkan's best preset
Dextracting the lora then gives you a similar results?
Fextracted lora should give better than lora itself
Flora training
DI swear I'm getting better likeness, much quicker with 2 repeats. no overcooking
Pit seems like my W11 machine keeps going into engery savings mode while computing, any good idea on how to stop that, besides turning off the energy savings mechanisms?
CProbably noon question. Please don't be mad. How I can add grid prompts to SwarmUI and use it?.
Or for example. I have trained my own face. Have 9 different loras. Now I want choose best. I'm doing all manually. Is any way to generate one Pic automatic using all loras and create a grid pictures?
Or for example. I have trained my own face. Have 9 different loras. Now I want choose best. I'm doing all manually. Is any way to generate one Pic automatic using all loras and create a grid pictures?
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dor 65 
Dcannot answer, but that's the reason I'm using forge
Fyes did you set to performance?
Fi have shown this
Fexactly
FYouTubeSECourses
Ultimate Kohya GUI FLUX LoRA training tutorial. This tutorial is product of non-stop 9 days research and training. I have trained over 73 FLUX LoRA models and analyzed all to prepare this tutorial video. The research still going on and hopefully the results will be significantly improved and latest configs and findings will be shared. Please wat...
F49:46 Overfitting in FLUX training and training image quality
51:59 Finding the best checkpoint using the Grid Generator tool in Swarm UI
52:55 Grid Generator tool for selecting LoRAs and prompts
53:59 Generating the grid and expected results
56:57 Analyzing grid results in Swarm UI
57:56 Finding the best LoRA checkpoint based on grid results
51:59 Finding the best checkpoint using the Grid Generator tool in Swarm UI
52:55 Grid Generator tool for selecting LoRAs and prompts
53:59 Generating the grid and expected results
56:57 Analyzing grid results in Swarm UI
57:56 Finding the best LoRA checkpoint based on grid results
Fi am gonna share so much information even full logs :d
D@Dr. Furkan Gözükara Now flux lora training is much slower, before I achieved 4.8 - 4.5 it/s, now I'm around 8.5. why is that?
Fwho knows need full debug on computer :d
Dbut you experienced the same right?
Dis something on kohya?
DIt happened and then was fixed with torch 2.5
Fnope i didnt
Ftorch 2.5 speeds up
Fyou need to run properly to no return back 2.4.1
Dso I should have the same speed 4.8 - 4.5 it/s on lora training with the current kohya version?
D
DI'm updating torch
Dmay be i was on an older version again
Dyour speed up script have pip uninstall xformers --yes
Dwhy?
Dxformers don't work with torch 2.5?
VAnyone else notice their flux fine tunes drop in likeness/quality really fast the second the subject gets far away from the camera/small? Like close up headshots look great, but a picture of the subject 20 feet away their face not only stops looking like them but starts being glitchy
VAny way to fix this?
D@Dr. Furkan Gözükara I updated torch to 2.5 but kohya installs 2.4.1 again when starting
DI fixed it --noverify
EHi everyone. Maybe this isn't news to many of you but just wanted to share my experience. I finally got fed up with windows using about 1.5gb of vram. So yesterday I bit the bullet and installed Linux, first I tried Mint but just couldn't get the Nvidia drivers to work so I switched to Pop_OS. It was a battle and so many bash commands later finally got comfyui and kohya working. And OMG! In win11 for Flux Dev Q8 gguf 50 steps at 1024 it took about 1:30 per image generation, that's after loading the models another minute+. But in Linux I'm now getting image gen in 32 seconds and models also load in just a few seconds. I'm using a 3090, 12700k, 64gb . I just wanted to share because I can't belive the difference. If you haven't tried SD with Linux and have the patience to get things working it's definitely something worth looking into.
Dthat speed is slow for windows, the speed should be similar
V@Dr. Furkan Gözükara I’m doing a flux Dora with the same dataset as the lora I showed you, quality is 1000x better.
DDora is suported in kohya?
VOnetrainer is what I’m using but I read that kohya scripts supports Dora, I can’t remember if that was a fork or implemented. I read it on either one of the kohya GitHubs.
Dthanks
D@Dr. Furkan Gözükara do you know that parameter function? I'm training flux-dev-de-distill and this model have a optimal cfg scale of 3.5, I was testing leaving it at 1 but it worked just fine.
