Fi think looks ok
Fthose could be related to preview generation perhaps?
CThere is alot to learn from the analysis of their graphs Dr. Furkan
C
CProbably something worth learning to use in order to optimize models
Fwell i havent seen anyone very correctly able to extract information from graphs. at best i know that lower loss is better 
CI am running dreambooth now because I want to see if I can replicate Alia checkpoint
CIf I can replicate a model to be identical, then we can scientifically analyze dreambooth data
Fthere are also some randomization on each run
Fespecially very erratic for lora i found
CThere is no such thing as random in computer science
Cnot if you get theoretical math, if I understand it correctly 
Cits not random, but unstable process if I understand it right?
SSir, I was able to train 2 subjects yesterday with 4500 epochs. I also used at least 1000 class images. The results were not very satisfactory. Is it beneficial if I continue training?
Fthere is no true random very correct
Fmore training usually dont get improved
Cmy model was at 22.600 and my hypothesis is that you could theoretically train to infinite given the right tokens
C75 tokens have a lower weight than 1 token
Ccompounded
Fdefinitely
Fthis thing has
Fbetween relationships as well
Fthat is not something
Fthat can be understood
Fmanually
Fit is all about machine learning
Fbigger scheme it works
Cyes
Fwhen you zoom in
Fall you see is bunch of numbers
Cyeah
Fso when you use more prompts during training
Fmore relationships are
Fcompounded
CHow many good images do you have of yourself Dr.Furkan, that are different?
SThank you 
Clike "epic" pictures of yourself
COk, but output
Fye just understood
Fi got quite good
Fafk for a while
Cok
So pick 300 quite good images that are unique and different.
create a model where you use the original 12 images that you used
12 images PLUS the 300 images WITH filewords.
This should not overtrain the model and because of compounded weight it should DECREASE the variation of the loss/average learning rate
So pick 300 quite good images that are unique and different.
create a model where you use the original 12 images that you used
12 images PLUS the 300 images WITH filewords.
This should not overtrain the model and because of compounded weight it should DECREASE the variation of the loss/average learning rate
CYou compound the most at 75 Tokens
CSo theoretically if you scan 300 images with 1 Token and 300 images with 75 tokens the Loss average/ learning rate should go down
CThis I will test now
CBecause actually I think there is replicated results even though there might be some standard deviation currently in the process
NHi again, i followed the instructions of the video. I found them easy although i dont know anything about code. Now, i am not getting the black samples during training but only the sanity sample is related to the subject. The prompt sample is random
Cblack samples?
Cwhat model are you running?