making same query for lora now
making same query for lora now
KThanks, It'll be helpful. I just read your comment on someone's query on the lora video, that it's not possible on 4gb Vram, wish i had read earlier. haha, great scratching my head, learned a lot though.
Fye 4 gb not possible :/
Iwhat is the significance of the loss value during lora training?
Iis a lower value more desirable?
Iim finding that i need to redo training 3 or 4x before getting a good result
Ilike restoring from the previous state, not fresh every time
Iso a total of like 15-20k steps
Flower loss is desired but doesnt mean well training aone
Fit can memorize and have 0 loss
Flora vs dreambooth pretty close quality
F
Fbut dreambooth has better background flexibility
Fi used 256 rank for lora
VHello
Is this Google colab still working with dreambooth?
https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb
After I've installed the dreambooth extension, when I try to restart all I receive this:
ModuleNotFoundError: No module named 'discord_webhook'
Is this Google colab still working with dreambooth?
https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb
After I've installed the dreambooth extension, when I try to restart all I receive this:
ModuleNotFoundError: No module named 'discord_webhook'
Fi think google colab banned it
Vso run pod is the solution?
Fyep
F
Fi am training a nice voice to do voice over and see how well it be on next video
Fit says that voice itself is not copyrightable :d
H256 rank？what is that
Iwhat does it mean when your lora is only good at drawing the face up close? Further away, like full body shot you dont really see the face at all, it more matches the base face of the model
Iis it because you didnt have enough far away shots or does it just need more training?
Iit affects how large the file ends up being, i think 256 would be like 300mb
Iif i want better results, should i be setting my rank to 256? i just do 128 like your last lora tutorial
Flora network size
Fi think go 256
Iwill that fix the face not working if its not close up?
Iif i have 12-15 pics, is 40 reps, 600 reg images 1 rep, 12 epochs for like 8000k steps enough?
Ii see people with much better loras on civitAI that clean to only used 2-3k steps, i am training at like 15k and not getting as good results
Iim wondering why my loras dont work very flexibly, is it because i am not labelling the training set?
Ii can prompt just for the character and get good close up results, but if i try to prompt a scene the lora basically appears to stop working
Iwould doubling the learning rate help?
IDo you lower the repeats the more images you train with and retain the same quality?
RHey just one thing. Dr doesn't use captions in the tut, but here you mention having captions, but you basically just use the command to add your instance prompt to the caption right? So you see a benefit to having a caption text files for Subject Lora's? That's the only thing that didn't seem clear to me
Fi tested captions today and they didnt improve for me once again
Ftraining certain subject much different than those models on civitai
Iwow i get much better result after only 3 epochs doubling the learning rates to 0.002 and increasing rank to 256
I0.0002*
Ianyone else have a bug where the first time you start to train it wont see the regularization images ?
Ioh for some reason the copy folder data didnt do reg image folder
Ihow did you say I could speed up training by increasing batch size or batches?
Iive been sharing your tutorials in all the SD groups im in
Ican you just keep training a lora and it will get better? or can you overbake them like embeddings?
KI put 1500 images in the regularization folder and SD still try to create more. Later I found out that some of my images are not 512x512 so I resized them and everything work fine. Also don't forget to put "/" before your directory path if you are using Runpod (such as: /workspace/stable-diffusion-webui/test)
Ihe said you dont have to resize images to train a face lora
Ihigher resolution better
KOh is that so? Did he mean training images or regularization images?
KAll methods can be overcooked. And in my experience it usually is around 150-200 epochs
Ihow can people have so many epochs? i do 4k steps in 3 epochs is this wrong?