[Stable Diffusion] Prompt Sharing and Learning Thread

[me3]

When the AI decides to completely ignore prompting for "real" and just does it's own thing...
I guess a working title would be "Puss.....in boots joins Star Wars"


Continuing the long running image generation (long running in a literal sense).
One image takes 1,5 hours to do, reaching the point where artists could probably do a very good job by hand in the same amount of time...

SDXL​	You must be registered to see the links	You must be registered to see the links
​	​	​
You must be registered to see the links (not upscaled)​	You must be registered to see the links (not upscaled)​	You must be registered to see the links (not upscaled)​
​	​	​

Spoiler: prompt

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

(Edited to add images from different models with same prompt and seed)
(Second edit, the last three images aren't upscaled to drastically reduce time, i also found out that the workflow/details wasn't getting save. If anyone is interested in it, any of the last 3 images should have it, one of the advantages of only having changed the model, so it's at least accessible/recoverable)

 

[BKKake666]

anyone have a bestiality lora model to bless me with?

pixiv has one... add this after domain: /en/artworks/105885164 (i cant post urls here yet)

 

[me3]

Just to spam more boring images ppl probably don't care much about. There's boobs this time though...
Did some tests to see how the amount of steps affect the image in XL, and a small comparison on models to see how a person would look.
Done in a1111...but thanks to the absolutely amazing optimizing there i had to run it with --lowvram due to it having absolutely horrible memory spikes and very weird times. So yes you can use XL in a1111 on very low specs, just used one of the models not requiring a refiner and have something else to keep you busy while waiting.

Images generated using dreamshaper xl:

No idea what's with the 10 step one

Model grid using 30 steps, links to the models can be found in this post:

Spoiler: prompt

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

If you look at the prompt you might notice there's one specific thing in there that all the images failed to include, but hilariously the cat in the linked post took that instruction very well

 

[Artiour]

Just to spam more boring images ppl probably don't care much about. There's boobs this time though...
Did some tests to see how the amount of steps affect the image in XL, and a small comparison on models to see how a person would look.
Done in a1111...but thanks to the absolutely amazing optimizing there i had to run it with --lowvram due to it having absolutely horrible memory spikes and very weird times. So yes you can use XL in a1111 on very low specs, just used one of the models not requiring a refiner and have something else to keep you busy while waiting.

Images generated using dreamshaper xl:
View attachment 2873940
No idea what's with the 10 step one

Model grid using 30 steps, links to the models can be found in this post:
View attachment 2873939

Spoiler: prompt

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

If you look at the prompt you might notice there's one specific thing in there that all the images failed to include, but hilariously the cat in the linked post took that instruction very well

try two handed sword (just a suggestion, I don't know the outcome)
as for the cat, how about this, you don't write a thing about any cat and yet there is one in the picture anyways

Spoiler

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

the mic and the stage/lights ...etc probably came from the tag platform heels, and the drooling of the cat maybe from "dripping", but the cat itself, I would so much love to make that cat[/SPOILER]

 

[me3]

try two handed sword (just a suggestion, I don't know the outcome)
as for the cat, how about this, you don't write a thing about any cat and yet there is one in the picture anyways
View attachment 2874143
the mic and the stage/lights ...etc probably came from the tag platform heels, and the drooling of the cat maybe from "dripping", but the cat itself, I would so much love to make that cat

You're using word(s) that are synonyms of it, and the model might be in some way biased towards that type of image. It's easy to notice background bias in models if you render alot of images without specifying things that requires a specific type of background.

And no, models generally don't understand "simple" concepts like two handed swords. or naming types of swords. You might have something close to it show up, but it's not held at all or just in some random way. One of the things you use controlnet for.

 

[Sepheyer]

 

[Sepheyer]

Space-MILF fever.

Spoiler: More

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

 

[Jimwalrus]

Space-MILF fever.

View attachment 2877559

Didn't know they had camels in space...

 

[Sharinel]

Didn't know they had camels in space...

miniature Giant Space Camels. Go for the lips Boo!

 

[me3]

Since XL is meant to be "better" at some things i wondered how it would be to create images in some XL model and then train a lora using those on a 1.5 model. If it worked it would if nothing else be a way to have "new" faces pretty easily as, lets call them default, faces in XL will be different than the ones we're very familiar with from 1.5.
So i generated a bunch of 1024x1024 images, since even that would be pushing my ability to train on them, dropped the ones with obvious problems/faults and started training...and noticed the expected time (stopwatch replaced with calendar)
Anyway, ran it for 8 epochs, sample images suggested the output was pretty consistent, so i stopped training and tested the loras and they showed a similarly pretty consistent "look".

This is 4 images generated on the trained model,

You must be registered to see the links

:


Did a test across some other models to see how it much of a difference there would be:


Is there any interest in the lora?
I've got no idea what kind of issues it has with regards to flexibility etc, i just did a very basic training setup and since it worked seemingly ok there wasn't no point in doing anything more.

(Updated to add link etc)

Adding image for advertisement?

You must be registered to see the links

, please no sale/profit usage, no claiming credit etc, just the usual respecting other ppls work

 

[Mr-Fox]

Since XL is meant to be "better" at some things i wondered how it would be to create images in some XL model and then train a lora using those on a 1.5 model. If it worked it would if nothing else be a way to have "new" faces pretty easily as, lets call them default, faces in XL will be different than the ones we're very familiar with from 1.5.
So i generated a bunch of 1024x1024 images, since even that would be pushing my ability to train on them, dropped the ones with obvious problems/faults and started training...and noticed the expected time (stopwatch replaced with calendar)
Anyway, ran it for 8 epochs, sample images suggested the output was pretty consistent, so i stopped training and tested the loras and they showed a similarly pretty consistent "look".

This is 4 images generated on the trained model, cyber realistic v3.3:
View attachment 2878282

Did a test across some other models to see how it much of a difference there would be:
View attachment 2878283

Is there any interest in the lora?
I've got no idea what kind of issues it has with regards to flexibility etc, i just did a very basic training setup and since it worked seemingly ok there wasn't no point in doing anything more.

yes ofc we want it.

 

[Mr-Fox]

Since XL is meant to be "better" at some things i wondered how it would be to create images in some XL model and then train a lora using those on a 1.5 model. If it worked it would if nothing else be a way to have "new" faces pretty easily as, lets call them default, faces in XL will be different than the ones we're very familiar with from 1.5.
So i generated a bunch of 1024x1024 images, since even that would be pushing my ability to train on them, dropped the ones with obvious problems/faults and started training...and noticed the expected time (stopwatch replaced with calendar)
Anyway, ran it for 8 epochs, sample images suggested the output was pretty consistent, so i stopped training and tested the loras and they showed a similarly pretty consistent "look".

This is 4 images generated on the trained model,

You must be registered to see the links

:
View attachment 2878282

Did a test across some other models to see how it much of a difference there would be:
View attachment 2878283

Is there any interest in the lora?
I've got no idea what kind of issues it has with regards to flexibility etc, i just did a very basic training setup and since it worked seemingly ok there wasn't no point in doing anything more.

(Updated to add link etc)
View attachment 2878762
Adding image for advertisement?

You must be registered to see the links

, please no sale/profit usage, no claiming credit etc, just the usual respecting other ppls work

Thank you very much, very generous.

 

[shkemba]

10 hours isn't as bad as you might think. My Kendra Lora took ap 16 hours with 3 epochs, 23 images, 768 res, 100 steps per image, 2300 steps in total per epoch, using slow learning rate and settings for "dampening".

You must be registered to see the links

Kendra Lora

You must be registered to see the links

View attachment 2816845

Can you please reupload the file? It seems to have been on Anonfiles (RIP). Thanks

 

[Mr-Fox]

Can you please reupload the file? It seems to have been on Anonfiles (RIP). Thanks

RIP indeed.

You must be registered to see the links

 

[LuminaJR]

just a few

 

[felldude]

A

You must be registered to see the links

for XLSD trained on images of the 4090 (If I had a 4090 maybe I could train at native res multiple epochs)
Although I did train the text enoder for this one and it did seem to help based of my experiments with clip

 

[Sepheyer]

Token Grouping vs Non-Grouping

So, (A),(B) are not the same as (A, B). Its kinda obvious, what's not obvious is where to use that in your prompts, i.e. what part of the rendering would benefit from such re-arrangement. In my case the model Zovya's RPGArtistTools wasn't sensitive to (middleaged) while other models were. I always wrote it off as a model's quirk. Then instead of having ~(woman)(middleaged) rewrote it as (woman, midleaged) and had the very response I was hoping for.

Went from this:

to this:

And here one more, although using a different model, with (oiled skin)(tan) vs (oiled skin, tan). Went from this:

to this:

 

[HardcoreCuddler]

A

You must be registered to see the links

for XLSD trained on images of the 4090 (If I had a 4090 maybe I could train at native res multiple epochs)
Although I did train the text enoder for this one and it did seem to help based of my experiments with clip


View attachment 2880075

sick design, actually.
Kinda offputting that this GPU has...are those USB ports? )) AI's are awesome

 

[KBAC]

How can this be fixed ?
OutOfMemoryError: CUDA out of memory. Tried to allocate 128.00 MiB (GPU 0; 4.00 GiB total capacity; 2.19 GiB already allocated; 0 bytes free; 2.30 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

 

[me3]

How can this be fixed ?
OutOfMemoryError: CUDA out of memory. Tried to allocate 128.00 MiB (GPU 0; 4.00 GiB total capacity; 2.19 GiB already allocated; 0 bytes free; 2.30 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Depends on when you get that message. If it's when you're loading the model (and you aren't already starting with this option), you need to add --lowvram to the command when launching.
If it's when generating images, you can try reducing the width/height of the image.

By the looks of it you got a 4gb graphics card but it's only able to allocate 2.3gb, so there's something else running that's eating much of your it, you could try closing any other software that's running before launching and see if it can use more of the vram as well.