btw a 128/128 extracted Lora will be like a gb per piece, right?
btw a 128/128 extracted Lora will be like a gb per piece, right?
Fit totally depends rank and precision
Flet me check
F128 rank float precision is 2 gb
F128 rank bf16 would be 1 gb
F64 rank bf16 would be 500 mb
Dmost of the time I see it recommended that alpha is half of rank
Falpha is just multiplier of the LR
Fi think it wont matter in this case
DI'm curious, did you do experiments with training images with transparent pngs? no background at all, just purely the subject
Fi didnt for person
Falso tranmsparent becomes white or black
Fdepending on script i think
Fno transparent level exists in stable diffusion
Tactually there is a transparent concept in Forge with layer diffusion, but it is a layer calculation by generation, working great but impossible to import that kind of technique in a lora (for now)
T
Fi know there is 1 model
Fspecifically trained and designed for this
KHi @Furkan Gözükara SECourses I've messaged you on Patreon, and you've suggested enabling
xformers
Mgo here
Fthanks for reply before i did :d
Mno problem 
DI've seen mixed advice, what's the latest thoughts on xformers for lora/fine tuning
JHi all, a total noob to training here and just have experience with Automatic1111 on RunPod. Any advice on the best training option for realistic people using RunPod? Dreambooth, Kohya, OneTrainer etc?
K@Furkan Gözükara SECourses after running OneTrainer for ~6 hours, I was finally excited to see the results of my DreamBooth training. But when I try to load the model in A1111 WebUI, I'm getting this error:
I've trained using the slow preset, 100 epochs. Only change I've made (apart from folder locations) is to enable xformers. What am I doing wrong?
I've trained using the slow preset, 100 epochs. Only change I've made (apart from folder locations) is to enable xformers. What am I doing wrong?
Ftry to avoid if possible
Fbut recent versions are much better than previous
FKohya dreambooth on runpod
Fi think i replied your message on patreon?
Fdid you take any backups_?
Fcheckpoints?
Fwhat is your file size?
Fi think it couldnt understand your model class like sdxl or sd 1.5 but not sure
KYes, I’ve taken backups every 10 epochs, so let me see if those work.
File sizes are 6.7GB for backups and the final file
File sizes are 6.7GB for backups and the final file
Flooking good
Fye try backup
Jyes thanks!
KManaged to find a working backup at 90 epochs, but the images generated look nothing like the training images. I've used 33 images which I feel are of relatively good quality, and the 5,400 regularization images from your Patreon.
Should I be increasing the learning rate?
Should I be increasing the learning rate?
Gsir, when will u work on RVC or any other better voice cloning tool
Fcan you share screenshots of your concepts setups
Fit is on my list but these keep coming up new ai models giving no time 
Gon point sir
Gsir SUPIR is taking too much time on kaggle while video upscaling.. is there any solution to this?
Fsupir is ultra slow
Fyou need a good gpu to run
Gok sir thanks
SDr Furkan, i hope every thing is ok, still one trainer is better option that kohya for GPU like 4070 super? 16 vram?,
Fi think yes probably kohya will cause out of vram error
Bhey how do you use sample imagines in dreambooth? how should the dataset look like?