is there an ideal ratio of reg images to training images?
is there an ideal ratio of reg images to training images?
Factually the more you use it is better
Ftherefore we try to use max
Fbut as you do more repeat
Fit will do more epochs
F200 repeat actually means 200 epoch
Fnot exactly due to scripts random image training logic but almost equal
Zok so 130*39=5070, which is close to 5200 which is the number of regularization images. Is that number similarity a coincidence or is that what I'm aiming for, for the number of repeats multiplied by the number of images to be about equal to the number of reg images?
Fye i calculated that way
Fbut it could lead to overtraining
Fsince it will be equal to
F130 * 39 * 2 steps
Fusually 15 images with 150 repeat yields good results even little bit overtrained for realism
A
Fmust be out of vram
Fshow task manager
Fgpu
Fyou know i can connect your pc and fix your python installaating and everything with private lecture
For you can follow this tutorial and instrall everything properly https://youtu.be/-NjNy7afOQ0
YouTubeSECourses
Welcome to the Pillars of AI Application Libraries: How to Install Tutorial! 
Are you ready to embark on a journey of installing AI libraries and applications with ease? 
In this video, we'll guide you through the process of installing Python, Git, Visual Studio C++ Compile tools, and FFmpeg on your Windows 10 machine. 
We'll also show you ho...
Are you ready to embark on a journey of installing AI libraries and applications with ease? In this video, we'll guide you through the process of installing Python, Git, Visual Studio C++ Compile tools, and FFmpeg on your Windows 10 machine. We'll also show you ho...
ZWhat is the most effective method of inpainting now?
Fcan you eloborate more?
ZI see people making many crazy cursed images on Facebook. I try on stable diffusion and just doesn't work, like this one of krampus and pit bulls
ZI assume they are inpainting the details
Fyes mask the area in auto1111 and inpaint
Fi think they are doing that
ZDo you recommend inpaint models, controlnet inpainting or sdxl?
Fi think just masking and inpainting with good prompt should work for that level of image
Fbut you can try sd 1.5 controlnet inpainting nodes too
A
APastebin.pl is a website where you can store code/text online for a set period of time and share to anybody on earth
PNeed some help, need to ideate on what the background should look like. Is there something in control net I can use to ideate some background ideas?
Cc: @Dr. Furkan Gözükara
Cc: @Dr. Furkan Gözükara
DAnyone recommend a good lora dim/alpha for realistic sdxl models? Is 128/1 still best or has everyone moved to the dreamboat extract mode now?
Bfor one epoch?
Zif I'm correctly understanding what he's saying, an "epoch" is a "full run" of training on your data set. So multiple epochs constitutes multiple runs of training on your data set. This is functionally the same thing as doing repeats for images. so your total number of steps is [training images] x [repeats] x [epochs], and take that and double it if you have regularization images. So doing 15 training images with 150 repeats and 1 epoch is 2250 steps (or 4500 with reg images). If you instead did 15 training images with 10 repeats for 15 epochs, that would also be 2250 steps (4500 with regs).
Mguys I have a question. Why does it give me a different output every time I copy a prompt from civitai and paste it into sdxl with the same settings? Shouldn't he give the same? Do I have to install like texture inversion? Help me
L@Dr. Furkan Gözükara I am playing around with different settings and one model is giving me issues in training a Laura. Most go smooth based on your video. I've tried more pictures, less pictures, more repeats, etc. I'm trying a higher learning rate right now. My go to settings are usually 2 batches(4090 doesn't appear to handle 3 or more on AdamW) 15 repeats and 15 Epochs. Default learning rate. This has worked awesome with 14 images or I've had great results with 8 repeats, 15 epochs, default learning rates, but 29 images. My subject is a weightlifter and I'm training with SDXL Base. Any tips?
Fwhat are you trying right now?
Fsomething else is looking wrong
Fnot using vram
Fit need further testing
Fbut i tested lora extract
F128x128 and 192x192 and 256x256 almost same
Fwhen you extract
Fout of dreambooth model
Fyes. because 150 repeating making it 150 epoch actually
Fsince every training image is 150 times trained
Fseed matters
Mbut i put them
Fi noticed same. sdxl realistic vision was undertrained at my settings
Fcould have extra loras etc
LI used a Learning rate of .001, thew results were much better! .001 might be high but because I'm doing parody I don't really need exact I just need to capture their likeness and make them something else