Because too many variables could waste your time...
Because too many variables could waste your time...
Fi would say do a fresh install
Fdont install torch 2
Ftest model there
IOk, which version to revert to? I will do fresh install, and check every single thing + extensions to see what hppened.
Fi think just fresh install latest version of web ui
Fdont install anything else
Fput your model and yaml there
Ftry
IWith torch 2?
Fno
Fdefault let it install whatever is default
IOK. thanks. I will try and update you.
Syes but there is no guide on their parameters and stuff
Fye i plan a guide yet as i said still couldnt find to test :/
Fsome info about safe tensors : https://huggingface.co/docs/diffusers/v0.14.0/en/using-diffusers/using_safetensors#what-is-safetensors
ZYeah. Created s new model and it stopped. I canceled and restarted like eight times so maybe that caused something to go awry
ZWill be between the unit learning rate and the tenc learning rate?
ZThe ratio
SI am bad at this!
Fso you compared checkpoints?
Zwhat is weight decay?
SDo you mean the incremental images? It seems like the best sample image is epoch 15. Is that what you mean?
Zmy model is sexy af
SError running install.py for extension extensions\sd_dreambooth_extension.
Command: "B:\Apps\Stable Diffusion\stable-diffusion-webui\venv\Scripts\python.exe" "extensions\sd_dreambooth_extension\install.py"
im getting this error from dreambooth new update
Command: "B:\Apps\Stable Diffusion\stable-diffusion-webui\venv\Scripts\python.exe" "extensions\sd_dreambooth_extension\install.py"
im getting this error from dreambooth new update
TTry back to previous state maybe restore point. Personally i dont update sd extensio if im having good results from current settings
Sok ill try
TDreambooth still very unstable
BQuestion. I want to make talking avatars, but D-ID and Pipio are too expensive to make the investment worth it. 15 minutes worth of vid for several dollars? Is there a better or at least cheaper alternative?
GIt seems like I am over training my model with my face and I am not sure how I can prevent this ?
Ti have a question, if I want to train specific outfit, what class should i use in Class Prompt in Dreambooth?
TAnyone notice in generating images that nomatter how you replace your promts or even delete all text and replace new compostion, after web ui has been reset and still the composition and style always generates the same? something like your stuck on the same seed. How to totally get away with that?
TPersonaly i found a good results with the method of 16 images 768,High res, Lr 0.0001, constant, 100epoch.
GNice! I will give that a try
Gdoes anyone know where the sample imgaes get generated in RUnpod ?
GAlright so I am not sure if I am doing this right. My CFG scale needs to be really low for my face to generate and even then it's not looking great
GThis is with the cfg scale on 1 and that is creepy AF but it's all warped
GAlright. after some testing, the new version of Dreambooth seems to over train VERY quickly. 150 for the face seems way to much
Gwhen I go to checkpoint 2 it's already overtrained , no style and my face is not even correct . I do think I might be doing something wrong
Zmy model keeps getting tripped up and making pictures of the girls drink lmao
Z
EI trained my model in Textual Inversion on SD 2.1. Results were not that bad, but I decided to check SD 1.5 because I was able to do hypernetwork on sd 1.5 and not on 2.1. I used same images and filewords in 1.5 as I used in 2.1. My embeddings have comparable loss value every 500 steps. But it is learning bad. Samples are much worse than in 2.1. Any idea what options or anything that I can change to improve my results? I used 22 images maybe I need to use more?
Gthis is the closest I can get my face into any image and then the ears, eyes and mustache is wrong
GI look like a 50 year old here XD hahahah
GI don't even have a mustache 
GAll of this is on epoch 100 so I need to fine tune a little more . Not sure how
KHello, it's me again.
I trained the model based on a cup, and it gave me pretty good sample images, but when prompted after training it doesn't give me the result i wanted. I used ohwx mug as a instance prompt and photo of ohwx mug as a sample image prompt
What have i done wrong ? Any ideas how to get the result from the training ?
I trained the model based on a cup, and it gave me pretty good sample images, but when prompted after training it doesn't give me the result i wanted. I used ohwx mug as a instance prompt and photo of ohwx mug as a sample image prompt
What have i done wrong ? Any ideas how to get the result from the training ?
Gwhat checkpoint are you using ?
KThe ones nearest to the sample prompts in question.
KWhen Prompted in text2img (with the model at 3000 selected - due to best sanity result) with the keywords : photo of ohwx mug, it gives me random mug images. The gray mug is the image i used (total of 25 images), and it gave me pretty good results during training.
KMaybe the base model i used for training already used ohwx ? maybe that's the problem?
