are you going to put out a new video for Lora training? @Furkan Gözükara SECourses
are you going to put out a new video for Lora training? @Furkan Gözükara SECourses
Eselamlar stable diffision kullanmaya çalışıyorumda dreambooth ile yüz taraması yapmaya çalışyorum 4060 laptop versiyonu 8 g vram li ekran kartım var 4 5 gündür birçok yöntem denedim bir türlü beceremiyorum anladığım kadarıyla xformers kurmam gerekiyo ve ekran kartı mimarisinden dolayı otomatik kurulum yapamıyorum internetteki bi çok metodu denedim defalarca yeniden stable diffusion kurmak zorudna kaldım Memory Attention bölümünde xformersı göremedim dün gece bir yöntemle xformerı görmeyi başardım bu seferde "Exception training model: 'import of xformers halted; None in sys.modules'." hatası aldım birçok yöntem denediğim için kafam biraz çorba oldu açıkcsı sanırım çok hızlı güncellendiği için sistemler bulduğum yönlendirmeler düzgün çalışmıyor sanırım xformersmı kurmalıyım başka yöntemler var mıdır yardımcı olma imkanınız var mı acaba
Fhi welcome
Fnumber of epochs 100% depends on your training dataset
Ftherefore best approach is saving checkpoints and comparing after training ended
Fi have done 15+ trainings so far to find best settings
Dthat's great to hear, ill be waiting for your next video. I have learned a lot from your SD videos.
Zgetting this error while trying to run 4x ultrasharp
ZError completing request
_pickle.UnpicklingError: invalid load key, '\x08'.\
_pickle.UnpicklingError: invalid load key, '\x08'.\
Ztrying to get 4x ultrasharp to work in my automatic111, i don't even have the ultrasharp in the esrgan folder anymore, yet they still show up here -
AThanks a lot for your response. So I guess I will try 200 Epochs and go from there. I guess the same applies to the amount of classification images, or is there a reccomended amout or formula for those?
Zi can't even count how many times i've broken automatic1111 already.
Zbetween compatibilty issues, different UI's, upscalers and extras not working right, dreambooth errors
Zbut because of that, i can setup an auto1111 installation in linux and be traiing a mode in about 10-15 minutes
Fwell formula depends on your training set so best way is checkpoint testing with x/y/z
Fi am doing another experiment
Fbatch size 8 xformers 0.0.19 dreambooth style training
Fbatch size makes huge speed increase
Ftesting batch size 10 lets see if rtx 3090 can handle it
Fbatch size 10 failed
Fbatch size 7 , rtx 3090 dreambooth ema xformers lion - didnt cache latents:
Fwow cache latents make huge diff - 6 batch size
Fi am going to sleep hopefully see you tomorrow
Hadditional networks extension？what is that? A name for a extension ？
H
HDownload first
HBut didn’t work
HThanks a lot
Hsolved by loading lora from additional networks
DDo you know what batch size is optimal for 3060 12GB? I have successfully trained using 2 Batch on 3060.
XHFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name':
XAnyone got error like this while using the pretrained model on colab?
AGood morning! I have a question regarding training with Dreambooth. I used the same dataset of 180 pictures of myself and trained using sd_1.5 and Deliberate as the base, with the same parameters. However, for each case, I generated the class dataset (photo of man) with its corresponding model. The results are as follows: when using the model trained with SD1.5, my face doesn't come out very well. When using the model trained with Deliberate, my face still doesn't come out very well. However, when I do the following checkpoint merge, the result is really good: (Primary: Deliberate, Secondary: Custom_SD1.5, Tertiary: SD_1.5, Add Difference, M=0.5).
In other words, I add the difference between my model trained with SD_1.5 and the original 1.5 model to the model I want to use as a base (Deliberate), and this is much better than fine-tuning the Deliberate model with my dataset. Does anyone have any idea why this is happening?
In other words, I add the difference between my model trained with SD_1.5 and the original 1.5 model to the model I want to use as a base (Deliberate), and this is much better than fine-tuning the Deliberate model with my dataset. Does anyone have any idea why this is happening?
HD:\SD\stable-diffusion-webui\venv\lib\site-packages\torchvision\transforms\functional_tensor.py:5: UserWarning: The torchvision.transforms.functional_tensor module is deprecated in 0.15 and will be removed in 0.17. Please don't rely on it. You probably just need to use APIs in torchvision.transforms.functional or in torchvision.transforms.v2.functional.
Hwhat should i do
Fyou are welcome
F2 can be the max
Fare you fixing revision? like main
F180 pictures really a lot to teach a person
Fhopefully today i will release a new tutorial for kohya ss train
Frealistic photos
Fyou can ignore
HOK thanks
AThanks for the advice. I've reduced the dataset to 70 photos, I will continue to perform more tests while waiting for the new video tutorial
ABy the way, for all my tests I followed your guide 'RunPod Fix For DreamBooth & xFormers - How To Use Automatic1111 Web UI Stable Diffusion on RunPod', however, when the model is being trained, an Out of Memory error occurs when generating the sample. I tried the A4500 (20GB) and the GTX 3090 (24 GB), but the error still persisted (even though the training only used 18.1 GB of VRAM with both cards). In order to train, I had to set the 'Save previews frequency' field to 0.
AI will do that, thanks 
YHello, could anybody tell me if classification images need to be square for Dreambooth in automatic 1111?
Fare you killing previous instance? also there is memory leak so you should restart web ui instance after you get an error
Fi think they should be same resolution as your target desired resolution
Fbut probably not needed squared