[Stable Diffusion] Prompt Sharing and Learning Thread

[Synalon]

Very Basic guide on how to use Reactor

Spoiler: Steps

You don't have permission to view the spoiler content. Log in or register now.

 

[Halmes]

Hey. Whats the best model to use training lora?

 

[Synalon]

Hey. Whats the best model to use training lora?

That would depend on what checkpoint you intend to use it with.

If you want to use it mostly with checkpoints trained on SD1.5 then use SD1.5, if you use mostly XL checkpoints use SDXL.

If you only ever use 1 checkpoint then just train on that.

 

[Halmes]

Im using google colab to train. I have 735 pics. How many repeats and epochs should i do? 10 epochs, 3 batch, 10 repats says too much steps.

 

[devilkkw]

What are you training? Style or subject?
700 is really too many pics.
I'm speaking about sd1.5 . In mostly case i use max 20 image, 18 steps and 4 epoch, for subject, and 36 step and 3 epoch for style.
I'm unable to test sdxl, train sdxl lora on 6gb is impossible, or maybe i don't know the correct way.

 

[felldude]

Im using google colab to train. I have 735 pics. How many repeats and epochs should i do? 10 epochs, 3 batch, 10 repats says too much steps.

I would look at the mean ar error (without repeats) with that many images. If it is in the 100th decimal your probably ok
(I have locally trained high error lora's in the 10ths for ar error, I think I trained 680 images with an AR error of 0.4)

Train on the EMA checkpoint if you can for

You must be registered to see the links

(I can't find the un-pruned EMA checkpoint anymore)
The XL EMA weights I recall being 16GB I can not find the checkpoint anymore.

EMA is more important for refinement then for a LORA.
Unless your training a style, the base EMA model produces a lora with wider range.

You could do a batch size of 4 and probably 4 epochs as your well over 10,000 steps

 

[Sharinel]

PSA regarding Forge, looks like it is no longer being maintained. Pity, as it is my main program to generate images

You must be registered to see the links

 

[Halmes]

I would look at the mean ar error (without repeats) with that many images. If it is in the 100th decimal your probably ok
(I have locally trained high error lora's in the 10ths for ar error, I think I trained 680 images with an AR error of 0.4)

Train on the EMA checkpoint if you can for

You must be registered to see the links

(I can't find the un-pruned EMA checkpoint anymore)
The XL EMA weights I recall being 16GB I can not find the checkpoint anymore.

EMA is more important for refinement then for a LORA.
Unless your training a style, the base EMA model produces a lora with wider range.

You could do a batch size of 4 and probably 4 epochs as your well over 10,000 steps

Im training with this

You must be registered to see the links

 

[Halmes]

I would look at the mean ar error (without repeats) with that many images. If it is in the 100th decimal your probably ok
(I have locally trained high error lora's in the 10ths for ar error, I think I trained 680 images with an AR error of 0.4)

Train on the EMA checkpoint if you can for

You must be registered to see the links

(I can't find the un-pruned EMA checkpoint anymore)
The XL EMA weights I recall being 16GB I can not find the checkpoint anymore.

EMA is more important for refinement then for a LORA.
Unless your training a style, the base EMA model produces a lora with wider range.

You could do a batch size of 4 and probably 4 epochs as your well over 10,000 steps

Shold i change to the sd training model too?

 

[Halmes]

I would look at the mean ar error (without repeats) with that many images. If it is in the 100th decimal your probably ok
(I have locally trained high error lora's in the 10ths for ar error, I think I trained 680 images with an AR error of 0.4)

Train on the EMA checkpoint if you can for

You must be registered to see the links

(I can't find the un-pruned EMA checkpoint anymore)
The XL EMA weights I recall being 16GB I can not find the checkpoint anymore.

EMA is more important for refinement then for a LORA.
Unless your training a style, the base EMA model produces a lora with wider range.

You could do a batch size of 4 and probably 4 epochs as your well over 10,000 steps

Think the problem is the paywall in google colab

 

[felldude]

Think the problem is the paywall in google colab

Based off of the error message your trying to use CPU training and need to

You must be registered to see the links

The drop down your showing is non EMA base checkpoint so in that case I would just use what ever model is closest to the style your going for.

With an image count that high you might want to lower the LR to 1e-6 for the U-NET and 5e-7 for the TE

 

[Halmes]

Based off of the error message your trying to use CPU training and need to

You must be registered to see the links

The drop down your showing is non EMA base checkpoint so in that case I would just use what ever model is closest to the style your going for.

With an image count that high you might want to lower the LR to 1e-6 for the U-NET and 5e-7 for the TE

GPU connection cannot be established due to Colab usage
says this. So i have to pay to use gpu more.
I changed google account but now it says " An NVIDIA GPU may be present on this machine, but a CUDA-enabled jaxlib is not installed. Falling back to cpu. "

 

[Synalon]

PSA regarding Forge, looks like it is no longer being maintained. Pity, as it is my main program to generate images

You must be registered to see the links

There is a fork for it, but I haven't tried it yet.

 

[DrPepper808]

I just started playing with this, its been a lot of fun. I have three questions.
1) lets say I have a prompt in txt2img and I create 25 images, and one is Perfect the other 24 not so much. is there a way to take that one image and train SD to do a better job with the same prompt?
2) I used a Lora to lock in the face, but also need to lock in the cloths so I can try my hand at making a game. Is there a way to lock those in>
3) if I want a image without a background for ez import to renpy is there a prompt for that?

 

[devilkkw]

I just started playing with this, its been a lot of fun. I have three questions.
1) lets say I have a prompt in txt2img and I create 25 images, and one is Perfect the other 24 not so much. is there a way to take that one image and train SD to do a better job with the same prompt?
2) I used a Lora to lock in the face, but also need to lock in the cloths so I can try my hand at making a game. Is there a way to lock those in>
3) if I want a image without a background for ez import to renpy is there a prompt for that?

What UI are you using?

 

[DrPepper808]

What UI are you using?

You must be registered to see the links

... it was the first one I grabbed.

 

[devilkkw]

i've used a1111 for many time, but i've switched to CUI.
So some possible answer:
1) you can train Lora with result image, but result is not sure. Lora train is not so simple, understanding every value for training, and all settings is not so simple in the begins, but you can try and see. Experimenting is the way. For train lora i suggest to use

You must be registered to see the links

2) in the same way for face, you can use lora for dress.
3) not prompt, there are some extensions doing it, but i checked it only on CUI, i can't point to right extensions for a1111 because i never checked on it.

A1111 is good and i had a fun for many time, but CUI is more powerfull for me, that node UI permit some workflow that is so difficult to remake in a1111. But this is my opinion, there are many skilled people's and everyone work good on UI they use.

 

[felldude]

I trained a lora with AdamW8 Bit on Pony using Civitai, my 2048 trainings on base Pony kept failing so I used my custom model to train, unfortunately that is limited to 1024x1024 a size I can do locally but eh I'd rather not tie up my machine for 12 hours vs 10 minutes for SD 1.5

Without chubby lora



With Chubby Lora

 

[DrPepper808]

i've used a1111 for many time, but i've switched to CUI.
So some possible answer:
1) you can train Lora with result image, but result is not sure. Lora train is not so simple, understanding every value for training, and all settings is not so simple in the begins, but you can try and see. Experimenting is the way. For train lora i suggest to use

You must be registered to see the links

2) in the same way for face, you can use lora for dress.
3) not prompt, there are some extensions doing it, but i checked it only on CUI, i can't point to right extensions for a1111 because i never checked on it.

A1111 is good and i had a fun for many time, but CUI is more powerfull for me, that node UI permit some workflow that is so difficult to remake in a1111. But this is my opinion, there are many skilled people's and everyone work good on UI they use.

when you say CUI do you mean ComfyUI? I just started looking into ComfyUI, seems like what I was looking for.

 

[devilkkw]

Yes, CUI mean Comfyui.
Actually I'm not a PC, but if you start using it I suggest to add ComfyUi Manager. You find link for downloading by search it on Google.
If you like it, and want to try it I will share my workflow for removing background when I'm at my PC.