After 20 hours of training,… it says this. Please help
After 20 hours of training,… it says this. Please help
J
DIs this is sd 1.5
JOh I see no space on disk, should I restart all again, or can I continue from 7/8 epoch ??
Byou can continue, just use your last checkpoint (sixth one, I think seventh wasn't saved properly) as your base model (=instead of sdxl base model)
Bbut of course beforehand you need to free space
JYou mean I should use this instead of SDXL base 1.0 on Kohya training ?
Bcorrect
Bbut you need to adjust the settings, as the training will start from the beginning
JI have other problem… I added these trained 1 to 6 to the Lora folder and I can’t see them here on stable diffusion
Bso you need to adjust the number of epochs for training
Bbecause they are not loras, they are full models just like sdxl base model
JOk good to 2 I think so not?
Bso they need to be in another folder
JOhhhh ok
Byep
JAs checkpoint
JSo I should use the model here
Jand ohwx man
Jetc
Byup
MDid I mention that I love Prodigy for SD 1.5 Lora training?
I just tried it with a very poor quality image set. I extracted images from phone video, and CompreFace could only select 13 of them, almost all of which were a bit blurry. I started training this with 8x iterations on 64/64 with the usual Prodigy script with d_coef 0.6.
The end result could be said to be good, because the subject was similar, but produced images as blurred as the sample images. So it also captured the "style" of the sample photos well, which would be good, but in this case it's a drawback.
I tried using alpha 1, but by about the third epoch it was overdone. That would be fine if I wanted a Lora in a few minutes, but that wasn't the goal. So I reset the alpha to 32 and took the weight down to 32 as well. To slow down the overtraining, I also halved the d_coef to 0.3.
And the miracle happened! The result also yielded images with a score above 0.9 (it generated 0.97 the first time!), and it wasn't blurry! Fantastic! So for anyone interested in this script version, I've attached the file, and a sample prompt for you to experiment with.
This is not 100% "style"-free, but the quality is much better, and on 32/32!
So even after nearly 100 Lora, Prodigy continues to surprise and challenge.
I just tried it with a very poor quality image set. I extracted images from phone video, and CompreFace could only select 13 of them, almost all of which were a bit blurry. I started training this with 8x iterations on 64/64 with the usual Prodigy script with d_coef 0.6.
The end result could be said to be good, because the subject was similar, but produced images as blurred as the sample images. So it also captured the "style" of the sample photos well, which would be good, but in this case it's a drawback.
I tried using alpha 1, but by about the third epoch it was overdone. That would be fine if I wanted a Lora in a few minutes, but that wasn't the goal. So I reset the alpha to 32 and took the weight down to 32 as well. To slow down the overtraining, I also halved the d_coef to 0.3.
And the miracle happened! The result also yielded images with a score above 0.9 (it generated 0.97 the first time!), and it wasn't blurry! Fantastic! So for anyone interested in this script version, I've attached the file, and a sample prompt for you to experiment with.
This is not 100% "style"-free, but the quality is much better, and on 32/32!
So even after nearly 100 Lora, Prodigy continues to surprise and challenge.
II have a question. For some people, the results of SDXL Dreamboots look OK in general, but the quality of the image is very low: skin texture, eyes, eyelashes. What can be the reason?
JWich one should I use here for big realism?
MI 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