uuuh- 12gb aree not enough😫
uuuh- 12gb aree not enough
F@Furkan Gözükara SECourses you should explore this!
FLoRA: The classic. You can use it in your webui no problem.
LoCon: Has more learning layers, it is reportedly good at artstyles.
LoHa: Has more layers and new mathematical algorithms. Takes much longer but can learn complex things such as styles and characters at the same time.
LoCon: Has more learning layers, it is reportedly good at artstyles.
LoHa: Has more layers and new mathematical algorithms. Takes much longer but can learn complex things such as styles and characters at the same time.Fmaybe it helps for your animation video, but also for another comparision video
F12 gb is good for dreambooth
Fyes web ui supporting that
Factually they gonna make even further improvement
Fi am waiting that for video
Fcurrently working on animation video with dreambooth
Fhaving consistency + strong style is the hardest thing to achieve
Nyeah, i know- we share the same grafikcard 
but using lion seemed to be to much for the card. i did not use xformers as well.
but using lion seemed to be to much for the card. i did not use xformers as well.
Fwell we have to use xformers
Fno other option
Fbatch processing frames taking time
Nis there a missunderstanding?
Fyes you have to use xformers for 12gb vram
Nright now i do not use it- with the 3060..
Fwow it works?
Fthey must have improved memory usage
For you are using torch 2
Frecently i have been testing on runpod :d
Npython: 3.10.6  •  torch: 2.0.0+cu118  •  xformers: 0.0.17+b6be33a.d20230315  •  gradio: 3.16.2  •  commit: a9fed7c3  •  checkpoint: a2ac9a57f3
i will compare the results to the last training once it is finished. same input images and class images
i will compare the results to the last training once it is finished. same input images and class images
Ftorch 2 has default xformers
Fi think that is the key
Nwhat is the best way while finding the best checkpoint using x/y-plot (checkpoint on x-axis, steps on y-axis) to use 4 completely different prompts for each example?
Fbest way is find a good prompt
Fget 8 decent looking seed
Fthen use x/y/z plot seed + checkpoints
Nit would be even better to have 4 different prompts to verify the results in different respects
Fyou can also do that
Fyou can add prompt with s/r option
Finto the x/y/z
Nbut not exactly what i ment- i really like (and do it the same way) your way to have a prompt with a batch size of 4 and only one prompt- the advantage is to do 4 at a time, but the disadvantage is that they are very similar- and i thought that i can define single parts of the prompt different with s/r but not the wholw prompt
Nsks as a superhero by zack snyder
sks as a cute pixar character
sks as a painting of van gogh on a 3-legged-horse
srs as a pencil drawing
sks as a cute pixar character
sks as a painting of van gogh on a 3-legged-horse
srs as a pencil drawing
Nsth like that..
Fi understand. you can also add different prompts to the x/y/z plot with s/r feature
Fso you can compare checkpoints, prompts and seeds all at 1 run
Nit seems that i have to do some more tries- as i thought that this wouldnt work easily
Fman i hope the method i am planning eliminates flicker
N"oil painting of a cabin in the wood, art by olivier ledroit"
Prompt S/R: olivier ledroit, vincent van gogh
Prompt S/R: olivier ledroit, vincent van gogh
Ni will have to put my old 1660 into my server pc to use to instances for doing stuff at the same time...
N"prompt"
s/r: "sks as a superhero by zack snyder","sks as a cute pixar character","sks as a painting of van gogh on a 3-legged-horse","srs as a pencil drawing"
not sure if that works at all- as once you have a "," in the prompt it could fail. will try in 58%
s/r: "sks as a superhero by zack snyder","sks as a cute pixar character","sks as a painting of van gogh on a 3-legged-horse","srs as a pencil drawing"
not sure if that works at all- as once you have a "," in the prompt it could fail. will try in 58%
ZGot it!
Zim ecstatic
Fwow looks like perfect
Fyou made it?
Zyup
Fwhat was the problem
Zi did about 22 images flipped, turned out to be around 44 images, it was cause i was using instance and class tokens without filewords
Zi didn't even notice that whole section was for filewords
