[Stable Diffusion] Prompt Sharing and Learning Thread

[Mr-Fox]

You are probably right, I forgot about resolution, after all this time I still can't understand what is CFG and sampling steps doing to be honest xD

edit: the problem that I have is that when I start the picture with higher resolution, even with the same aspect ratio, sometimes I get split images, or a collage of images, so clearly there is something I'm missing

Yes when you use a higher sampling resolution, SD has a problem handling this and you get monstrosities.
Such as conjoined bodies etc. In rare cases you can get a split image. Make sure that you didn't forget to turn of batch count or batch size. If this is set to 4 you will generate 4 images and in the end an image grid or collage. Batch count is how many seed steps you want, batch size is how many images from the same seed you want.
Example:
If you use batchcount 2 and batchsize 4, you will use 2 different seeds and 4 images of each, 8 images in total.
This is a very useful tool to generate many images in the search for the perfect one.

 

[fr34ky]

Yes when you use a higher sampling resolution, SD has a problem handling this and you get monstrosities.
Such as conjoined bodies etc. In rare cases you can get a split image. Make sure that you didn't forget to turn of batch count or batch size. If this is set to 4 you will generate 4 images and in the end an image grid or collage. Batch count is how many seed steps you want, batch size is how many images from the same seed you want.
Example:
If you use batchcount 2 and batchsize 4, you will use 2 different seeds and 4 images of each, 8 images in total.
This is a very useful tool to generate many images in the search for the perfect one.

That's golden info for me, I never knew what the difference was, and never knew how to generate more than 1 image of the same seed.

 

[daddyCzapo]

Today i installed xformers, and oh boy, this shit bussin'. Went from "out of vram when generating 512x768 with x1.5 hires upscale after 7 minutes" to "ez 512x768 x2 hires upscale in 3 minutes". And tbh after some testing, i see no change in quality whatsoever. And finally i can generate some portrait mode pics

 

[Mr-Fox]

That's golden info for me, I never knew what the difference was, and never knew how to generate more than 1 image of the same seed.

It increases the vram usage though. So it's not an option if we are trying to use as high resolution as possible. It's a shame that it is generating 2 or more at the same time when using batch size. But there are good news. You can use xy plot script instead.
Example:Set seed as X axis and something else for Y axis that you just repeat as many times as you want images of each seed.

 

[fr34ky]

It increases the vram usage though. So it's not an option if we are trying to use as high resolution as possible. It's a shame that it is generating 2 or more at the same time when using batch size. But there are good news. You can use xy plot script instead.
Example:Set seed as X axis and something else for Y axis that you just repeat as many times as you want images of each seed.

I don't use hi res for txt2img, upscale in img2img phase so it's not a problem for me

xy plot started to give me errors like 1 month ago and there was no way to fix, and now I don't even want to think about it, I don't enjoy much using it anyways.

 

[Mr-Fox]

Today i installed xformers, and oh boy, this shit bussin'. Went from "out of vram when generating 512x768 with x1.5 hires upscale after 7 minutes" to "ez 512x768 x2 hires upscale in 3 minutes". And tbh after some testing, i see no change in quality whatsoever. And finally i can generate some portrait mode pics

Xformer doesn't change quality. The rumor is that it makes the image generation less predictive. In my experience, this is not a thing. If there is an increase of variation it's so small that you can't differentiate it from the normal variation. Meaning repeated generation of the same seed and same settings. Each image will be slightly different. But this will happen regardless if you use xformer or not. Imo more variation is not a bad thing regardless. Yes, do use xformer. Any boost in performance we can get is well needed.

 

[Mr-Fox]

I don't use hi res for txt2img, upscale in img2img phase so it's not a problem for me

xy plot started to give me errors like 1 month ago and there was no way to fix, and now I don't even want to think about it, I don't enjoy much using it anyways.

Oh man that sucks. I use it all the time, it's great. I run comparisons very often to weed out the best settings, wich sampler to use and many other scenarios. For instance I made several different iterations of my Lora that I could compare with the same prompt and seed thanks to plot script. I have never had any error specific to this. If nothing else works a simple re-install will likely fix it.

 

[fr34ky]

Oh man that sucks. I use it all the time, it's great. I run comparisons very often to weed out the best settings, wich sampler to use and many other scenarios. For instance I made several different iterations of my Lora that I could compare with the same prompt and seed thanks to plot script. I have never had any error specific to this. If nothing else works a simple re-install will likely fix it.

The error I get is AssertionError: VAE from function argument doesn't exist: auto

I have 2 other installations of SD and the error happens in all of them.

I used it in the past for some tests, but after this problem happened and it drained my brain energy I stopped caring.

 

[devilkkw]

Try this link .

You must be registered to see the links

Thank you, this is a test with your lora from 0.1 to 0.8 with simply prompt

Spoiler: simply prompt text

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

For me best value is 0.6, so try added some photographic trigger at these value. This is the result.


If it is your first lora, wow man, you rule!!

 

[fr34ky]

Oh man that sucks. I use it all the time, it's great. I run comparisons very often to weed out the best settings, wich sampler to use and many other scenarios. For instance I made several different iterations of my Lora that I could compare with the same prompt and seed thanks to plot script. I have never had any error specific to this. If nothing else works a simple re-install will likely fix it.

I can't confirm 100% but after some tries I could use x/y/z without problems, so I just set the z in 'nothing'.

Kind of weird, but if it works who cares, I hope it reminds like that.

 

[Mr-Fox]

Thank you, this is a test with your lora from 0.1 to 0.8 with simply prompt

Spoiler: simply prompt text

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

For me best value is 0.6, so try added some photographic trigger at these value. This is the result.
View attachment 2499646

If it is your first lora, wow man, you rule!!

Thank you very much. I'm pretty happy with it myself.
Try add "vivid light". You can scale it back if it's too much, "(vivid light:0.6)".

 

[fr34ky]

Any idea of what is the limit for words on the prompts? I'm searching about that and there is no clear answer, but there is something about 150 or 75 tokens and I've seen some long ass prompts that won't fit in anyway into that limit.

 

[Mr-Fox]

Any idea of what is the limit for words on the prompts? I'm searching about that and there is no clear answer, but there is something about 150 or 75 tokens and I've seen some long ass prompts that won't fit in anyway into that limit.

There isn't a limit but it is divided into groups of 75. This is the explaination I have seen. "A token is a sequence of characters that represents a single unit of meaning in a text. It is a fundamental concept in NLP, as most NLP models operate on a token level, meaning they process text one token at a time".

(NLP: Natural language processing in artificial intelligence ).

 

[Mr-Fox]

Come to bed..

Spoiler:

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

Yup, hands are still a big problem for SD. I hope this area improves rapidly.

 

[Sepheyer]

Spoiler: CUI Prompt

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

 

[Mr-Fox]

View attachment 2504321

Spoiler: CUI Prompt

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

Wow. That's one of your best yet.

 

[Jimwalrus]

Not sure if this has already been covered or not, but there is an extension in WebUI (Embedding Inspector) to allow weighted merging of embeddings, models etc.
It seems to be a great way to create very consistent characters for, ooh, let's say, GAMES without it being obvious that the character is just a merge of two celebs...

For instance, I merged my TIs of

You must be registered to see the links

and

You must be registered to see the links

at 0.5 each to create an entirely new embedding (I just called it SK1NNY, for fairly obvious reasons).

One part Kylie:

and one part Zoe Voss:

gives this young lady:


The influence of both is fairly clear, but only when you know the two source embeddings.

I ran some more of her as checks for consistency and they turned out pretty well.

Spoiler: More

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

 

[Jimwalrus]

Not sure if this has already been covered or not, but there is an extension in WebUI (Embedding Inspector) to allow weighted merging of embeddings, models etc.
It seems to be a great way to create very consistent characters for, ooh, let's say, GAMES without it being obvious that the character is just a merge of two celebs...

For instance, I merged my TIs of

You must be registered to see the links

and

You must be registered to see the links

at 0.5 each to create an entirely new embedding (I just called it SK1NNY, for fairly obvious reasons).

One part Kylie:
View attachment 2506528
and one part Zoe Voss:
View attachment 2506529
gives this young lady:
View attachment 2506527

The influence of both is fairly clear, but only when you know the two source embeddings.

I ran some more of her as checks for consistency and they turned out pretty well.

Spoiler: More

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

Further findings.
It's not exactly predictable. It doesn't appear to be "taking a pinch of celeb A and a pinch of celeb B", it seems to just be a largely random, but consistent, merge. Also, if one embedding is strong, it can overpower the other.

For instance, this is

You must be registered to see the links

merged with

You must be registered to see the links

each at 0.5 multiplier:

Pretty much Cameron Diaz. It is a very strong embedding.

Changing the weighting to 0 multiplier for Cameron and 0.4 for Kylie produced a bit of a surprise:

As you can see, internally consistent but unpredictable. The afro came from nowhere!

I shall continue my research into producing pictures of random attractive women by mashing together two unrelated celebrities...

Next thing to check will be how it compares to in-prompt mixing of embeddings i.e. "KYL13M:0.5:C4M3R0ND14Z:0.5:1"

EDIT: Also, yay 1000 posts on this thread!

 

[sharlotte]

Unpredictable but delicious and pleasing to the eye! I had read on it but never tried it before.

I had an issue where I created monsters (Cthullhu would have run for the subaquatic hills) with double of everything, legs over body, really awful stuff. I went carefully over every slider, options, and it turns out I was generating my picture with too many pixels mostly in the height dept... Anything over 1154 pixels in height just could not cope with the models I was using.

As part of this small fact finding mission, I started using the 'realistic vision 2.0' from civitai. Great results, far less monstrosities, defects than before. A few samples as png so prompts included. I did not use any hires on these.

 

[KoutaruArts]

Anyone got the Amber Heard Lora From Civitai??, the author either got banned or delete big part of hes content if some does please share it.