I use DPM++ 2M Karras and DPM++ SDE Karras
I use DPM++ 2M Karras and DPM++ SDE Karras
JOk great
JDo you use ADETAILER?
MNo, I use Hires Fix only. I hate Adetailer. It never delivered the results I expected.
JYea I thought same
JAnd do you use negative prompts?
Bhow many repititions did you use? gonna try it out later
Bah in the config it is 10 epochs so I assume you meant by 8 iterations = 8 rep
MI never touch epoch, it's always 10, and 99% of the time the last one is always the best. That's why I like Prodigy with cosine training, because I don't have to search for the best version because it's the last one.
I always calculate my reps like this: 100/number of images, and rounded up
I always calculate my reps like this: 100/number of images, and rounded up
MThese data are personal. For style (as you can see from the above) I need to train a stronger model, but there is labelling, so it's absolutely up to me how strong a style I want.
MBut if I stick to flexibility in style, I don't go above 2 for tensorboard normalization. Then I can mix style with the uniqueness of other models. But if I'm doing a very strong style, then it almost doesn't matter which model I choose to use.
MI had 13 pictures, so 13*8 = 104
Bpretty good for sd 1.5. thanks for sharing your results
MThe great thing about Prodigy is that you can control the learning speed with two switches. The d_coef is the fine tuner, and the alpha is the coarser one. The lower the alpha, the faster it learns. So up to 5 epochs would be enough if the alpha is half the weight. This is just a theory, I haven't tested it, but it would make it even faster to train Lora, even in a few minutes (now about 10-15 minutes on an RTX 3090 card).
MI like the 10 epoch and the 100 divisor because it's more transparent, and based on previous tests, it's really the right one, no more, no less. And for that you can already set the d_coef well.
MAnd that's how I get the photo quality under SD 1.5, plus I can get a Lora model ready really quickly. No need to train DB.
MI haven't switched to RealisticVision 6.x for training yet, I think 5.1 is good enough to get Lora results that work well on other models, and I get nice results on RV 6.x.
MSo with these results I think it's unnecessary to train on SDXL for now, maybe I'll think about SDXL-turbo if I can train it to get even better results in a small size.
MAlthough someone recently argued for SD 2.1, that it could also be trained nicely, but many people quickly gave up on it. It's 768, so it might be worth a try with Prodigy.
MHello my kaggle account got blocked can i create a new account using the same phone number?
OGitHub
Features:
settings tab rework: add search field, add categories, split UI settings page into many
add altdiffusion-m18 support (#13364)
support inference with LyCORIS GLora networks (#13610)
add l...
settings tab rework: add search field, add categories, split UI settings page into many
add altdiffusion-m18 support (#13364)
support inference with LyCORIS GLora networks (#13610)
add l...
XIsn't that the new president of Argentina?
KI have a lora for lineart style, an image(simple lineart + colored lineart) both then how can we change the style of the little lineart only? Using controlnet or ipadapters? Or change the style of the plane colored image by some anime style lora?
MSomeone explain this to me in a simple, understandable way, what is it for in 1.7.0 version?
https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13568
https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13568
GitHub
Description
Some Trainer (like hcp-diffusion, naifu) have implemented a method of training called "pivtoal tuning", which is basically trained Embedding and LoRA(DreamBooth) at the same t...
Some Trainer (like hcp-diffusion, naifu) have implemented a method of training called "pivtoal tuning", which is basically trained Embedding and LoRA(DreamBooth) at the same t...
Pthe funny thing is that in the dev branch there are still 132 commits that are not in this release, a lot of things are still to come and be adjusted.
FIt say right there in the issue:
- We can now train Embeddings (aka Textual Inversion) and LoRA at the same time
- However, multi-concept models produce many embedding files, one for each concept.
- This adds a new feature to read "bundle" files that bundle many embeddings together into one
Fsadly i am not that expert
Fyour space full
Fyou can download existing checkpoints
Fbut it wont continue
Fyou need to calculate number of checkpoints
Fyes
Fi see this was on runpod
Fyou can delete older models
For turn off pod
Fexpand storafge
Fand use last checkpoint as a base model
Fand continue
Ffor delete use rm -r file_path
Fare you using adetailer to improve?
Fno you need a new number
Flooks like just another lora method
KSo.... Anyone else stuck in Check for Updates in A1111
KAnd counting
K
Fnewest update could have bug
FUDEMY banned my course and it is too much bothersome. i wonder if anyone among us could help me?
KWell... I removed some extensions I don't use anyway and now it works 
KBanned!?!?! Fu**ing why?
Know I remember why I used Roop and not Reactor
JDoes anyone know why the photos generated when the face is close up are perfect, and when it is shown further away, like with the whole body, it looks bad?