The Doc has a video on training a LORA but he never mentions an embed so maybe it's trash?
The Doc has a video on training a LORA but he never mentions an embed so maybe it's trash?
FI think that "Save your Model Frequency" Is a very important value, I did just that, the memory usage goes crazy because at the end of the training you have lots of models. I assumed as a noob that the more the code works on a model the better it gets but that´s not the case, I think its called overtraining so... I ended with 10 versions of a model and apparently the third one is working very good, the fourth one gives bad results the fifth one very bad results and so on...
FNo! realistic results!
Spfft, if that's what you're going for, I suppose. But it's blurry nonsense for me! LET'S GO!
FHo! no, I he mentions that LORA is more difficult so I´m starting with embed 1.5
F
Syour pic looks more like my subject than mine does
Sthat is dope
FPROMT: painting of p3p34gu1L4r by rembrandt, high resolution. 8 k 3 d render trending on art station 4k intricate artwork masterpiece digital painting concept illustration greg rutkowski beeple alphonse mucha cinematic lighting atmospheric light glow octane unreal engine photorealistic hyperrealism photography complex detailed matte print fantasy style atmosphere surreal portrait canon eos c 300 ƒ 1 5 0 mm f 2 - 6 sony alpha n
Fhahaha
FTry this:
I think that "Save your Model Frequency" Is a very important value, I did just that, the memory usage goes crazy because at the end of the training you have lots of models. I assumed as a noob that the more the code works on a model the better it gets but that´s not the case, I think its called overtraining so... I ended with 10 versions of a model and apparently the third one is working very good, the fourth one gives bad results the fifth one very bad results and so on...
I think that "Save your Model Frequency" Is a very important value, I did just that, the memory usage goes crazy because at the end of the training you have lots of models. I assumed as a noob that the more the code works on a model the better it gets but that´s not the case, I think its called overtraining so... I ended with 10 versions of a model and apparently the third one is working very good, the fourth one gives bad results the fifth one very bad results and so on...
SI am saving every 50 steps so i have like a million, each worse than the next!
S
Fdamn!
FI know how frustrating that is, believe me... I spent 4 days watching nightmarish distorted monsters until now
FHow many photos are you using?
S18
S19?
Fthats a good number I think
SI am sure my settings are wrong
Ssooo many buttons
FRemember that the photos should be in 512X512 and they need to be very different from each other, different angles, lenses, backgrounds, clothes etc
FYes! Don´t lose hope!
FI have follow this tutorial https://www.youtube.com/watch?v=Bdl-jWR3Ukc&t=3153s 4 times in the last 3 days, from beginning to end, every time a notice something new, so much to learn!
YouTubeSECourses - Software Engineering Courses
Our Discord : https://discord.gg/HbqgGaZVmr. The most advanced tutorial of Stable Diffusion Dreambooth Training. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on 
https://www.patreon.com/SECourses
Playlist of Stable Diffusion Tutorials, Automatic1111 and Google Colab ...
https://www.patreon.com/SECoursesPlaylist of Stable Diffusion Tutorials, Automatic1111 and Google Colab ...
Sis anyone experienced with lora training here?
FIs there some bug in A1111 dreambooth training where the checkpoints it generates just ignore all the training? it generates samples fine but then try to use the checkpoint to generate images and the model you generate works but just ignores all the training instance tokens completely
For resuming training is broken or something, not sure wth is going on
FI read somewhere that the latest Dreambooth is just broken, no matter what you do, what parameters you adjust the results would be nightmarish
It´s so frustrating! Spent days trying for a model to work and to understand the app but the results are always garbage after hours of waiting
It´s so frustrating! Spent days trying for a model to work and to understand the app but the results are always garbage after hours of waiting
Fhow do you mean garbage?
Fdistorted nightmarish
Fthats not exactly the issue i'm facing, it is training, its generating good samples, but when i try to use the checkpoints in SD to generate images, it ignores the training and just treats it like its the underlying model with no training at all.
Fits not doing a crappy warped version of what i ask it to do, its just ignoring the training altogether
EI got error like this while training hypernetwork:
"A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check."
I tried to add --no-half or set the "Upcast cross attention layer to float32" but it wont help. Wheni used --disable-nan-check then error disappeard BUT my samples are all black.
"A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check."
I tried to add --no-half or set the "Upcast cross attention layer to float32" but it wont help. Wheni used --disable-nan-check then error disappeard BUT my samples are all black.
Fi did a new train yesterday it worked for me. i didnt generate any previews later did x/y/z
Flatest dreambooth working. i just did train yesterday
Fwhat model you are training sd 2.1?
Eyes
Fyou need to add --no-half
Fthat requires more vram
EYes i know 
I then I have error that I dont have vram 
I thought maybe I can resolve this in another way
I then I have error that I dont have vram I thought maybe I can resolve this in another wayE1.5 requires less vram?
Fyes 1.5 much lesser vram
Ffor example i cant do dreambooth on sd 2.1
Feven on 512 px
EI thought maybe it is because I resized the images from 768 to 512 using tool in automatic1111 and it somehow broke the images but i doubt that is the issue. Ok, I will try 1.5 then, thanks
THello Sd trainiacs. Can somebody share the command line to reset sd repo?
Fwhich version of sd 2.1 you are using 512 or 768
Fwhat you mean by reset?
Fyou can do git pull