Training 50 images with captions
Training 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
Fonetrainer way better
Fcurrently with flux we are literally cooking the u-net
Fi trained with only ohwx word
Fdidnt use in prompt only man
Fit still generates my image :d
Fnew test results published
F
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
SIs this with the new clip_l?
SOr unet only
Fstill unet only
Fclip l just arrived like 1 hour ago
Fi am waiting bmaltais
SOhh yeah ofc
Fi learnt the reason
Fflux has internal text encoding system that combines
Fgiven prompts + images
Fbut we are overfitting atm
Fi think fine tuning will fix it
SYeah I hope the clip_l training is even effective though
SThe lora extraction in kohya for flux is a bit broken sadly, results r poor and checkpoint itself is huge > 20GB