[Stable Diffusion] Prompt Sharing and Learning Thread

[devilkkw]

I hope you weren't aiming for accuracy in the "conversion". Even if we ignore all the small details that is a pain at the best of times with AI generation, the large "missing" bits would strongly suggest there's work to be done.
IE, breast size, even in a less anime nature they are clearly way too small and "flat". As for the clothing, neck area has completely changed, abdomen area has its clothing removed and instead you have gotten "rendering discoloring" which looks like what you get from overtraining. Crotch area is completely different as well.
It might be due to an unfortunate choice of picture and/or model, but if that's the case it wasn't the smartest choice of marketing example

I know, and it don't keep so much details because denoise is high, for keeping detail i have to keep slow denoise. and on really flat image is not good. maybe working on different way but actually i'm unable to train lora, so my only way is Textual Inversion. Btw is a good start point to work better on it, i like constructive comment, thank you.

 

[me3]

I know, and it don't keep so much details because denoise is high, for keeping detail i have to keep slow denoise. and on really flat image is not good. maybe working on different way but actually i'm unable to train lora, so my only way is Textual Inversion. Btw is a good start point to work better on it, i like constructive comment, thank you.

if it's a vram concern, you shouldn't need that much more to train loras than TIs, i can train on 768x768 images with just a old 6gb card which works fine. It's slow but it works, it should be possible at even lower vram too.
Should give lycoris a try, depending on type you can create files almost as small as TIs, about 200kb, unlike loras that is much larger.
You can also adjust lycoris better in the prompts to better work out training issues and, eventually, better fit to the image you are converting.

 

[Mr-Fox]

You and "both" the AIs did a good job, but those two images happen to perfectly illustrate a problem i've been having as well.
Getting the AI to do (what we at least consider) basic things like "carry/hold item in hand" can be a real nightmare, and even if you get it sort of right it's usually in a very unnatural way. If the seed itself doesn't include it, you're looking at an endless battle. Having spent nearly 60 hours trying to get someone to hold a sword in a way that was even remotely looked natural, i'd rather look for a seed where it was included on its own and then "fight" to add the things that was "wrong" in that new seed...

You either need to train a Lora, TI's etc or use open pose with controllnet. Alternatively find an already existing Lora or TI's etc, but this is always with the hazard of getting a poorly trained one.

 

[Mr-Fox]

I know, and it don't keep so much details because denoise is high, for keeping detail i have to keep slow denoise. and on really flat image is not good. maybe working on different way but actually i'm unable to train lora, so my only way is Textual Inversion. Btw is a good start point to work better on it, i like constructive comment, thank you.

If you can do checkpoints and TI's you should definitely be able to train Lora's. I used kohya ss for mine, don't try to use Auto1111 for this.

You must be registered to see the links

(copy the settings of 59xed3 instead of Aitrepeneur).

Here's a very extensive guide about the actual training:

You must be registered to see the links

Also take a look at Schlongborn's awesome post about training his first Lora:
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10306324

 

[me3]

You either need to train a Lora, TI's etc or use open pose with controllnet.

Yeah i know, i'm still just very much in the "figure out what this thing can natively do" stage. I'm also of the kind that likes the challenge of trying to find ways/tricks/workarounds for things. It's also rather funny seeing all the "super amazing" lora/ti/models ppl say they've made to do stuff that models already know how. Having to resist the urge to ask/post "did you even test if the AI understood this first"

Currently down to about 15-20min training time on my current testing set, testing out which settings has any effect and "what". (4.8 s/it on an old worndown card). Might get that further down as it might be slightly overtrained, but i need to keep it constant for the moment to work out what other values/settings to use and hopefully be able to transfer much of that to other datasets. Problem with having such a slow setup is that by the time I have any meaningful results there's an update or a much better "something" and it's back to the beginning.

 

[devilkkw]

If you can do checkpoints and TI's you should definitely be able to train Lora's. I used kohya ss for mine, don't try to use Auto1111 for this.

You must be registered to see the links

(copy the settings of 59xed3 instead of Aitrepeneur).

Here's a very extensive guide about the actual training:

You must be registered to see the links

Also take a look at Schlongborn's awesome post about training his first Lora:
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10306324

I tested it time ago, by your older comment. i was unable to run it. maybe i see is update and giving it a new chance. thank for useful link.

 

[Mr-Fox]

I tested it time ago, by your older comment. i was unable to run it. maybe i see is update and giving it a new chance. thank for useful link.

I had problems at first also, I had to download "alternative cuda libraries" for my Pascal GPU and it solved it. I made a post about it. You probably needs something similar to make it work.
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10357013

 

[devilkkw]

I had problems at first also, I had to download "alternative cuda libraries" for my Pascal GPU and it solved it. I made a post about it. You probably needs something similar to make it work.
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-10357013

Thank you so much, new version work and i'm able tu run it, now is time to try and make some test.

 

[PandaRepublic]

Does anyone have a tutorial on how to make a LyCORIS model?

 

[Mr-Fox]

Does anyone have a tutorial on how to make a LyCORIS model?

I'm also interested in this. If you track down any good info source please share. With a search I found these links:

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

 

[me3]

In a basic sense it's not much different from training TI or Lora. Most of the difference is related to the optimizer and not really the "type".

I have found some odd behavior in training recently and i'm not sure what is causing it.
Since i've been testing alot of different values/settings i created a "base" file which i reloaded each time to reset things incase i'd accidentally update a setting etc. I ran multiple trainings on that base file to see that the results were the same each time, generated images had just the slight random nature you'd expect like strains of hair being slightly different etc.
After quite a while of testing i started seeing results that didn't make much sense so i retrained on just the base file and images generated was nowhere close to what they initially was. Character was suddenly 3x as old and at least twice the body weight. Have spent over a day now trying to get back to the starting point i have no idea what's "gone wrong". Nothing updated, training data has remained unchanged etc so somewhat confused.

Another thing i noticed is related to epoch settings.
With the base file i had the training thankfully only needed 1 epoch, IE the -000001 file generated using kohya-ss, but i had left the setting to run for 5 epochs just incases. When i changed this setting without changing anything else, the images generated on the trained file (still using the first epoch file) changed as well.
So i did multiple trainings with setting the "max epoch" to different values then generating images with the same seed using the 000001 file from each of those trainings.
So keep in mind, these image sets are ALL generated on using the 000001 file from trainings done on the exact same settings, same data etc, only thing that's changed is how many epochs the training would have run for. That is the number in the group titles, so where it say "ep_2" means it was set to run for 2 epochs, but it was still the first epoch file used to generated images.


This was repeatable too, so running the same generating on the same seed gave the same results so it isn't due to randomness in the generating. For some seeds the difference was much greater as well, IE one image would have just the clothing and not even the person, while wearing the same clothing in another.

I'm not sure if this is due to randomness in the optimzier (Prodigy) or if it's related to the scheduler (cosine) or if it's some kind of bug in the code.
It could be that the "cosine wave", to call it that, gets stretched from start to end, so that more epochs affects growth/decay, no idea tbh. Because it takes so long to test it i can't really look into it.
Regardless of the cause the result for "us" is that it can affect our training results and it's a value you'd not expect to do that besides when things stop.

 

[Mr-Fox]

In a basic sense it's not much different from training TI or Lora. Most of the difference is related to the optimizer and not really the "type".

I have found some odd behavior in training recently and i'm not sure what is causing it.
Since i've been testing alot of different values/settings i created a "base" file which i reloaded each time to reset things incase i'd accidentally update a setting etc. I ran multiple trainings on that base file to see that the results were the same each time, generated images had just the slight random nature you'd expect like strains of hair being slightly different etc.
After quite a while of testing i started seeing results that didn't make much sense so i retrained on just the base file and images generated was nowhere close to what they initially was. Character was suddenly 3x as old and at least twice the body weight. Have spent over a day now trying to get back to the starting point i have no idea what's "gone wrong". Nothing updated, training data has remained unchanged etc so somewhat confused.

Another thing i noticed is related to epoch settings.
With the base file i had the training thankfully only needed 1 epoch, IE the -000001 file generated using kohya-ss, but i had left the setting to run for 5 epochs just incases. When i changed this setting without changing anything else, the images generated on the trained file (still using the first epoch file) changed as well.
So i did multiple trainings with setting the "max epoch" to different values then generating images with the same seed using the 000001 file from each of those trainings.
So keep in mind, these image sets are ALL generated on using the 000001 file from trainings done on the exact same settings, same data etc, only thing that's changed is how many epochs the training would have run for. That is the number in the group titles, so where it say "ep_2" means it was set to run for 2 epochs, but it was still the first epoch file used to generated images.
View attachment 2746744

This was repeatable too, so running the same generating on the same seed gave the same results so it isn't due to randomness in the generating. For some seeds the difference was much greater as well, IE one image would have just the clothing and not even the person, while wearing the same clothing in another.

I'm not sure if this is due to randomness in the optimzier (Prodigy) or if it's related to the scheduler (cosine) or if it's some kind of bug in the code.
It could be that the "cosine wave", to call it that, gets stretched from start to end, so that more epochs affects growth/decay, no idea tbh. Because it takes so long to test it i can't really look into it.
Regardless of the cause the result for "us" is that it can affect our training results and it's a value you'd not expect to do that besides when things stop.

This was very useful information. Thank you!
The more we can reduce variables, the more control we'll have over the training and the end result.

 

[phraskito]

View attachment 2730286
A lovely snow white, fit for a prince!

Please, What model have you used?

 

[Mr-Fox]

Please, What model have you used?

cyberrealistic_v30. It's very easy to find out with PNG Info in SD Auto1111.

 

[FreakyHokage]

Does anyone have a tutorial on how to make a LyCORIS model?

You need the extension to use LyCORIS. Go here

You must be registered to see the links

click "code" and copy the link then go to the extensions tab on automatic1111 and click "Install from URL" paste the link under " URL for extension's git repository " and click apply and wait for it to install then click reload UI. All LyCORIS models go into your LoRA folder.

 

[sharlotte]

Please, What model have you used?

Mr Fox is correct. Usually I post the png with info, so if you run in 'PNG info' in SD tab, you'll be able to see all the prompts and settings used on my generated images. Here it is

You must be registered to see the links

There's a newer version since 3.1.

 

[maikijo]

Tell me, please, why is it so? I write in prompts "one person" or "1 person", but in the end, a lot of people still appear in the picture.

 

[Dagg0th]

Tell me, please, why is it so? I write in prompts "one person" or "1 person", but in the end, a lot of people still appear in the picture.

View attachment 2753073 View attachment 2753074

Your images are not valid or are edited, so I can't check what are your prompts to see what are you doing.

One probable cause is that you are using different width and height, try a 512x768

Put "1girl, solo", in positive, use weight if necessary, (1girl:1.5) for example.

 

[me3]

I was gonna post a grid showing of image repetition and epochs, just incase someone found it useful to just see how it changes.
I'd done 10 runs where each run had an increase in how many times the training images was repeated per epoch and did 10 epochs each.
Even compressing it down to jpg won't let me post it before it becomes too distorted (it's only 11k x 10k pix as well), so instead i'll break the thread rules and just post a one of the images from the trainings that i "lost" the file for and been unable to recreate. Prompt wouldn't help even me as the lycoris file is long gone so
Still, looks like a pretty nice boat/yacht though, bit out of focus

 

[Jimwalrus]

I was gonna post a grid showing of image repetition and epochs, just incase someone found it useful to just see how it changes.
I'd done 10 runs where each run had an increase in how many times the training images was repeated per epoch and did 10 epochs each.
Even compressing it down to jpg won't let me post it before it becomes too distorted (it's only 11k x 10k pix as well), so instead i'll break the thread rules and just post a one of the images from the trainings that i "lost" the file for and been unable to recreate. Prompt wouldn't help even me as the lycoris file is long gone so
Still, looks like a pretty nice boat/yacht though, bit out of focus
View attachment 2754060

What boat?
Oh, the motorboat...