Then I am confused because I thought that was what you were asking about.
Then I am confused because I thought that was what you were asking about.
Fi am trying a deep fake tutorial :d
AI am trying to build a docker image for RunPod with WebUI V1.3.0 and latest main branch for Dreambooth that has the correct torch and xformers etc built in, and also all requirements matching so nobody has to mess around with everything and struggle to get things working. Also changing relauncher.py to launcher.py so nobody needs to mess with that either.
Fthe setting about "noise scheduling" where you can set DDIM, DEIS or UniPC, refers to how the noise is handled for the training. aka: generating noise to train the UNET, right?
AYes, but don't understand your question with regards to the classification images.
Fwe're generating the classification images with euler_a, right?
Ffor example when you download the patreon class images, those were generated with euler_a
AYes, correct
Fso we are telling the noise predictor to use DDIM noise scheduling, to train/predict the noise of the images
Fand we're using images generated with euler_a ?
Fisn't that weird?
AI could be wrong, but my assumption is that the class images as well as the training images are converted to noise with the same noise scheduler for training, so it is irrelevant how the class images were generated, but thats just an assumption.
生OK THX A LOT
Fexacly. so why use generated images that generate low quality stuff?
Ffor example, "photo of a man" generate this:
F
Fwhen we train our training images into the model, we're also training this class images !
Fwhy are we training the model with this stuff?
Fwe're asking the model to understand that "man" means "monster"
Fwhy class images are not photos of real people from the internet?
Fwe are teaching the model what "man" means
Fis not better to tell the model to learn that a "man" means this?
F
F
Fi am going to gym 
ALooks like Keanu Reeves 
FYep working on it
FI should publish a
Video today
Video today
FWhat other artist you suggest?
FCurrently at the gym
FMaybe terminator arnold
FFThe Dreambooth implementations out there typically introduce generated images out of the model itself into the training set. I do not suggest that, and that's why EveryDream is not a Dreambooth implementation. I instead suggest you can gather other alternative "ground truth" data and "mix it in" with your training data to "preserve" the model and avoid "forgetting" or other artifacts. The tools repo has a Laion data scraper, or you can use things datasets like FFHQ photo set.
I recommend "ground truth" images as opposed to the Dreambooth implementations which use generated images out of the model itself for preservation. Dreambooth training on generated images can reinforce the error already present in the model. Ex. 6 fingers, bad limbs, etc.
Fthis is what i've been saying!
Fvery interesting
Fgonna try it right now!
Fi think dreambooth extension without classification images does what everdreamer trainer does but i havent tested
D@Furkan Gözükara SECourses read my dm's
Feverdreamer uses classification images. but not generated ones
AExcuse my ignorance, but what is everdreamer?
F
Fthis guys says that reg images are useless
Fand captions is better
Fthe opposite of what people like Furkan says
Fthis is so confusing!! xD
Fit's another method for fine-tuning, like kohya_ss, etc
Eso this was it. i have new xformers now, and now my training in automatic 1111 dreambooth extension is working, but only in the sample images dreambooth folder shows me. , when i go to text to image, it is not recognizing my lora input at all. is there a known issue with stable diffusion and lora? theres always something 
(kohya trained lora seems to be working, I may have to just default to kohya instead of dreambooth extension )
(kohya trained lora seems to be working, I may have to just default to kohya instead of dreambooth extension )
MValueError: height and width must be > 0
It happens when I do with Inpaint. Why ?
It happens when I do with Inpaint. Why ?
Fi'm liking everdreamer, it's easy, you just spawn pod and execute a simple command
Fand results looks good, but i'm still on epoch 20, let's see