Yes I watched that video but you were showing the different upscalers there. I'll look again
Yes I watched that video but you were showing the different upscalers there. I'll look again
Fjust do small test
Flike 2 lora
F2 prompt
Fand see results
Fthat will give you idea
Falso play with very top options
�Ok thank you @Dr. Furkan Gözükara I think I'm figuring out now. At first I was just using prompt settings from top left only but now I have the prompt on second tab and it's getting much more closer to what I was going for
�Sorry I get so easily frustrated, I wanted to figure it out on my own and not have to bother asking you until I had no other option. Very much appreciated, Dr! 
Fi got to sleep hoepfully see you tomorrow
�Goodnight Dr. I just ended up using this workflow instead in the Comfy tab: https://civitai.com/models/652483/flux-lora-tester
Hello 
for whatever reason you want to show me appreciation, you can: 
Ko-Fi With this workflow you just save time. A lot of time. Have fun. To...
for whatever reason you want to show me appreciation, you can: Ko-Fi With this workflow you just save time. A lot of time. Have fun. To...
�I completely had forgotten how much VRAM it was going to take to load up 40 Loras at once 
but this works good for now
but this works good for now
KTesting loras must be easier in Forge
P@Dr. Furkan Gözükara for kohya flux Lora training, should it save a safetensor at each epoch? As it hasn’t been saving them in my model folder for some reason. It says “epoch is incremented”, which I haven’t seen before. Pretty sure in the past with Sdxl Lora training, it says “epoch saved”
P
PAlso it’s just finished training and has provided just one safetensor
Fi find swarmui best
Fmulti gpu?
Fyou training 880 images or what?
Fclip l arrived i am gonna train
P1 GPU
PTraining 50 images with captions
Fok you have setup error
Fmake folder repeat 1
Ftrain up to like 100 epochs at least
Fso it would be 5000 steps
Fsave every 10 epoch
PHmm, no I haven’t received an error.
I’m using your rank 4 config at the moment
I’m using your rank 4 config at the moment
PIt’s strange as I tried the tank 4 config on 20 training images last night, and it made around 170,000 steps which surely isn’t correct?
Fnot error folder setup
PIt reached only 8% training in 12 hours
Fyour folder setup inaccurate
Fmake repeat 1
Fi, plan to start recording tuıtorial video today hopefully
Fpay attention to folder names
PAh so here should be 1, not 40?
Fas a folder path i give this : R:\mass_compute\training_imgs
Fyes
P
Fmake 1
PIs this the reason it was making 170,000 steps?
Fyep
Fit makes 40x more
PThank you, I’ll adjust this
PI’ve noticed training flux on Renaissance painting details is polishing the result (giving a slightly plastic feel), not picking up the granular textures of paint, crack etc. Do you know why is this?
Is there parameters such as learning rate which will effect this or perhaps it’s the architecture of flux itself?
When training the same images on kohya Sdxl, the trained model will clearly be show the texture of painting and canvas surface etc.
Is there parameters such as learning rate which will effect this or perhaps it’s the architecture of flux itself?
When training the same images on kohya Sdxl, the trained model will clearly be show the texture of painting and canvas surface etc.
KI still think Kohya with it's folder-preparation and such feels like a stone-age tool :/
Fwe already use lower LR
Fwhat will impact is
Ftraining clip l text encoder - which arrived today
Fand full fine tuning
