i got a new finding
i got a new finding
Fshift value 1
Vi thought it was common to use shift value 1
Dstarting a training with TE
Vi've been using it this whole time
F
Fi was using default value 3
Falso this is shift value sampling
Dlet's see if still bleeds......
Fyou mean discrete_flow_shift right?
F
Fcolors better
Farmor better too
VOverfitted still
F
Falthough likeliness a bit reduced
Druns slower on WSL with TE
Mdid you upgrade to torch 2.4.0?
Dyes I did 2 days ago, I was having 3.5 to 4 it/s
Dnow I enabled the text encoder and it runs much slower
Myou training 1024?
Dyes
Dsomething is on cpu
DI think is T5 that is not training, just CLIP_L
Dbut really idk
Dmay be the performance can be improved
Dwhat about the speed training the text encoder? is running slower for me
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Ftime shift sampler comparison
F
Fconfigs added to folder
S@Dr. Furkan Gözükara I am having a OOM error in Forge, on Massed Compute, do you have any idea?
SEven when I lower this slider
Fcan you edit this file and restart let me tell you
Fbut before that actually can you dm me your entire screen screenshot
Fso i can see your exact config
Fbecause today i tested i have no error
SSure thing
D@Dr. Furkan Gözükara Is it possible to integrate joycaption as the default captioner for Supir, instead of llava?
Nrunning a big grid right now comparing different versions of a person lora... 25, 50, and 75 epoch versions all have long hair for "man as a cosmic jazz improviser whose music causes spontaneous formation of new elements and life forms and wearing eyeglasses" even though short hair in all training images 
Fi may integrate
DSince you've mastered custom LoRA training for people, I encourage you to explore Flux LoRA training for products, objects, and art styles.
Fyep it is also in progress
FI agree, I don't understand why I'm losing the comic style I trained my lora with via comfyUI. I do have my avatar but in a realistic style, whereas I keep the comic side when I go through automatic 1111. I'd like to train a style to have consistency.
DFlux is clearly excelling at art style and product LoRAs, as evidenced by the recent surge of these types on Civitai. Also I look forward to seeing architectural and other categories too.
Fi added new config
FBest_Configs_Better_Colors
Fit may be fix for style training
Fi will do more research with it not compoleted yet
F@NicB we got error with forge