but when i question the ans will show error
but when i question the ans will show error
生
Fwhat do you see on cmd
生
生it seem like my vram is too small?
Fif training uses DDIM or DEIS noise scheduling... why regularization images are generated using euler_a ?
Fshouldn't be generated using ddim?
Di have 1gb vram lol
Fand i still don't understand how to do regularization images!
Fsome of my images are like this:
F
Fwhy i would want to re-train my model with this? it makes no sense!
DLOL
DWTH IS THAT
DCURSED IMAGE
Dthats a doll mate
Fthat's a "photo of a woman"
Dwhat is that lipstickkkkkk 
Fthe class images stuff is very unknown!
Fye out of memory
Ftry smallest
Fit is what model already knows
Fthat gives flexbility
Fit can generate 500m images with unet :d
ADDIM and DEIS are not the only noise scheduling options, there is also UniPC.
And the Image Generation Scheduler options are:
DDIM
DDPM
PNDM
LMSDiscrete
EulerDiscrete
HeunDiscrete
EulerAncestralDiscrete
DPMSolverMultistep
DPMSolverSinglestep
KDPM2Discrete
KDPM2AncestralDiscrete
DEISMultistepUniPCMultistep
UniPCMultistep
So it is not a 1 to 1 mapping.
And the Image Generation Scheduler options are:
DDIM
DDPM
PNDM
LMSDiscrete
EulerDiscrete
HeunDiscrete
EulerAncestralDiscrete
DPMSolverMultistep
DPMSolverSinglestep
KDPM2Discrete
KDPM2AncestralDiscrete
DEISMultistepUniPCMultistep
UniPCMultistep
So it is not a 1 to 1 mapping.
Fbut image generation scheduler is only for class images
Fis not for training
Fis for generation
AFor class images and also for sample images.
Fexactly
Fthat's pure inference, has nothing to do with training
AThen 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?