Hi @Dr. Furkan Gözükara or Fellow AI enthusiasts. What is the best learning track for Creating Certa
Hi @Dr. Furkan Gözükara or Fellow AI enthusiasts.
What is the best learning track for Creating Certain images based on DataSheet and Input images?
I have Series of columns (Eg 5 columns in excel, those are desired image attributes) and Some Images.
I should get desired accurate output (Have few samples as well),
I have tried Lora models in stable diffusion, not sure onhow to feed datatbale as input, Some one please point right example or Tutorials for this
16 Replies
Hi guys, I just wanted to ask you something. Why does it give me an error on xl every time I try to add 1.5 embeddings? PS: I'm using kaggle
Hello. I want to ask you.
I maked lora on kaggle with tutorial video.
Then, I can't the file on pokemonpets.
Please tell me how to solve this.
Come on my DM
Hello. With 40 images and 100 repeats, without regularization images training steps is 4000 and near 4000 steps it becomes to get overtrained with my set of images. So today I wanted to try using regularization images but i forgot that when you use reglarization images the training step count becomes X2 so in my case 8000 steps. So my question is since without regularization images around 4000 steps the results become overtrained, with regularization images and training steps up to 8000, wouldn't it still get overtrained past 4000 steps? And if the answer is yes, should I half my repeats setting to 50 to get 4000 steps or should I keep repeats setting at 100 and just download the checkpoints created until 4000 steps? Thank you.
Are you Lora training?
no dreambooth
might extract lora later
I don't have much experience with DB but I find 4k steps to be high. I shoot for around 3k. Lowering the repeats is probably the best move
thank you 👍
@mikemenders im quite new to understanding tensorboard, what is it I need to look out for during training to see whether the LORA is going to come out good or not?
If you use my training script, it has the advantage of using the cosine method of normalization. In this case you have to look at two graphs on the tensorboard, the rest is irrelevant. One is max_norm/keys_scaled (on left side) and the other is max_norm/max_key_norm (on right side).
On the former you can see how strong Lora is. If it only scores 1, it's a bit weak, but good. If it scores 2, it's good, it takes a little style from the images (pose, quality) but it will still be flexible. If it reaches 3 once or twice, it's borderline, a stronger Lora is made, which is harder to format. Above 3, I prefer to let styles go, the point of which is to make it strong enough to push through models.
The max_norm/max_key_norm is relevant because of where it reaches the ceiling, i.e. 1. Usually it is good to reach it after a third of the training, but before half. Because after half of the training it is more of a grinding of the model, no major modification is done, because we are training with cosine method. So in my experience, it's better to reach 1 before the halfway point. As soon as it reaches it, it starts to normalize and spikes appear on the max_norm/keys_scaled graph.
Here is a cool graph of a good Lora training. This Lora generated images with a score of 0.97. Here the value of d_coef was 0.75.
Here are two poor graphs of two Loras. In the case of the pink one (on the right), you can see that it doesn't reach 1, so it's under-trained, it won't be good. The yellow one reaches it, but the normalization rarely works, so just a good Lora, it will not be too strong, i.e. it will not always generate good images, only rarely. And yellow reaches 1 halfway through the training, so there is not much time left for a stable Lora.
amazing thank you, I just trained one but it didnt come out very good
By default, I can already see on the sample pictures whether it will be a strong lora or not, because it already resembles the person at the third epoch, but if it's already at the second, it will definitely be too strong. Then halfway through the training you can see on the graph how strong it will be. If it already makes two 3s, I stop the training, lower the d_coef and start again.
For you, the graph on the right is good, but the one on the left... Strange. So it reaches 1 there, but you zoomed in? Because if it reaches 1 on the right, it should show at least 1 on the left, if not more.
I’ll try trying again, I’m getting strange results at the moment, faces look straight and no details (skin ect)
*strange
This is partly fixed by Hires Fix, and of course depends on your source images. If they don't have much detail because they're a bit blurry, Lora will take care of that too. But this is something that Hires Fix helps with (depending on the base model).
here is a generated photo in 512:
and with Hires Fix:
