[Stable Diffusion] Prompt Sharing and Learning Thread

Thalies

New Member
Sep 24, 2017
13
50
So, I’m pretty new to all this, but I just had to share. I made my first image using SDXL Turbo in Fooocus. Probably not a big deal for most here, but for a noob like me, it’s pretty awesome how fast it is and how good the quality is.

You don't have permission to view the spoiler content. Log in or register now.
 
Last edited:
  • Like
Reactions: Sepheyer

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
SDXL turbo is not only experimental. You can find many ckpt's by simply use the filter on civitai "SDXL turbo" or simply search for SDXL Turbo. For SD1.5 there is LCM, that is essentially the same thing. High quality results with much fewer steps and a lower cfg scale. LCM: " Latent Consistency Model" .
SDXL Turbo.png
SD15 lcm.png
 
Last edited:

hkennereth

Member
Mar 3, 2019
237
775
I tried it now, it does still persist. It only works when I start with a seed right away on all samplers.

But when I create a new random seed that I like, and I liked the outputs, there is no chance to only re-run the 2nd or 3rd sampler, it would always start from scratch. Probably because I change the first sampler from "random" to "fixed".

But if I don't do that, I would get new seeds all the time. And if I have to put in the seed first and set it to fixed, I can't keep working with a first randomized generation that I like, that I now want to "lock in".
Okay, so if I understand it right, you are leaving the first sampler at Randomized so you can find a new image you like, and then when you find it, you are changing the seed to Fixed, is that correct? And that's why it's not working for you?

If that's the case, the reason is that the randomization of seeds happens AFTER you add the image to the queue (you can see that your primitive node, where it says "CONTROL_AFTER_GENERATE"). Meaning you are looking at a seed value of "5", for example, you click Queue Prompt, that image with seed "5" is added to the list, and then a new seed value "42" is displayed, so if you switch NOW to Fixed, you are using the value for the NEXT image, not the one you just created. A quick workaround for that is once you find an image you like, you can drag it back to the ComfyUI window, which will load all the values as they were for the creation of this image, including the seed, and THEN switch all samplers to Fixed, which will keep them at that specific value. Now you can fine tune all sampler values, prompt tags, etc, and they will all remain at the original seeds.

That Global Seed node I mentioned, on the other hand, has an option that doesn't exist on Comfy by default, which is the option to change the seed randomization mode to "CONTROL_BEFORE_GENERATE", which will randomize the seeds at the moment you click to Queue Prompt and then add the image to the queue, keeping all seed values as they were. This is how other UIs like A1111 and EasyDiffusion tend to work.
 
  • Like
Reactions: Fuchsschweif

Fuchsschweif

Well-Known Member
Sep 24, 2019
1,145
1,957
A quick workaround for that is once you find an image you like, you can drag it back to the ComfyUI window, which will load all the values as they were for the creation of this image, including the seed, and THEN switch all samplers to Fixed, which will keep them at that specific value. Now you can fine tune all sampler values, prompt tags, etc, and they will all remain at the original seeds.

That Global Seed node I mentioned, on the other hand, has an option that doesn't exist on Comfy by default, which is the option to change the seed randomization mode to "CONTROL_BEFORE_GENERATE", which will randomize the seeds at the moment you click to Queue Prompt and then add the image to the queue, keeping all seed values as they were. This is how other UIs like A1111 and EasyDiffusion tend to work.
Thanks, that are two great tips and the background information about how the seeds work was very helpful! I will try them both out. I already got the inspire pack installed too.

Edit: Works very well with the Global Seed node. :)
 
Last edited:

me3

Member
Dec 31, 2016
316
708
And I never denied that it could be my hardware or system, I just said overheating won't be the issue, which was something you guys seemed to hang up on. Also I don't know why you ask me to be more specific about my crashes, only to tell me that I basically should go troubleshoot the other 50 possibilites. It sounded like you were onto something and wanted to help out.



Dude, I got that from the very beginning, but you seem not to understand my approach at troubleshooting.

It was never about SD being bugged or the VAE decoder being the cause itself. It was simply about figuring out what the tiled VAE decoder does, that the VAE decoder does not do, so that I could at least roughly narrow down the direction I have to look into.

Anyways, I made another test. Since shutdowns never happened with the tiled decoder yet, even with larger upscales than those causing a shutdown, I tried to go even bigger (with the tiled version) to stress my system further and see if I can cause a shutdown.

And it actually happened, with a very big upscale size that goes far beyond the higher things that only work with the tiled version for me, but not with the standard ones.

So now I think the most likely reason it shuts down has to do with the PSU. This theory didn't make sense before, because shutdowns never happened during the heavy demanding upscalings but only in that short 3 second time-frame of conversion, but since I could trigger a shutdown now during the upscaling process for the first time, it points at running out of power or some instability in the PSU.

If it was only the GPU being at its limit I would suspect for it to just throttle down or the display shutting off. Overheating I did rule out with measurements yesterday. Faulty communication on the mainboard between components would probably lead to a bluescreen crash, since conflicting protocols would cause a software error.

Unless someone of you comes up with any objection, I tend to just go with a new PSU as it seems to be the most likely reason.
It might not be a "problem" with the PSU failing as such, more that there's spikes and very shifting fluctuations causing voltage "drops". There seems to be reported issues of things like tensorflow and pytorch causing reboots (without blue screens etc) and the fixes are mainly limiting the power usage for the GPU. Some refer to issues with bios versions but that obviously become much more specific to cards than a general power usage issue. One place at least mentioned ASUS x299 Sage having this bios issue. Not sure if that applies in your case or if similar issues exist with other cards.
You could try limiting the power usage for the GPU using nvidia-smi, one post i saw had to set power usage to 120 or less for their 1070 so might be a place to start. Probably a good idea to do a quick search and see what safe values are for the card.
Some had luck disabling a turbo mode for intel CPUs, but seems you are using AMD, might have something similar
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Hi everyone
I'm sorry if I bother you but I'm not English and therefore it's really difficult to search for threads.
Can you tell me if stable diffusion is still among the best or if it has been overtaken by XL or XL turbo or something else?

I like the idea of generating images in real time but I understand that the interface still works with the node system and differs from stable diffusion... I don't understand anything anymore

Can I please ask you what you suggest I use?
The main software is Stable Diffusion. There are mainly 2 user interfaces that is the most established. Automatic 1111 and ComfyUi but there are also others such as Fooocus (a simplifyed UI).
Regardless which user interface you choose, the software is still Stable Diffusion.
There are different model generations, just like there are windows 7, 10 and 11 etc. The ones that are used the most are SD1.5 and SDXL. I would recommend either going with Automatic1111 or Fooocus to get started, when you have become proficient with it, then try a less beginner friendly user interface such as ComfyUi (node based UI) if you wish.

How to install Stable Diffusion Automatic1111:
How to get started Using SD A1111:

Stable Diffusion Fooocus tutorial:

*Edit
Something to be aware of.
Fooocus is primarily using SDXL.
It is supposedly lighter on the vram requirement compared to A1111 though.
I don't know if you can download SD1.5 models and use with it.
 
Last edited:

Delambo

Newbie
Jan 10, 2018
99
89
Somewhere in here was linked a page that showed examples of all kinds of art styles by artist. I can't for the life of me find it anymore either here or in my history. Anyone got any suggestions that they think might be the one I'm thinking of, or a substitute for it?

Page had a blue background and broke out all kinds of styles by artist, then gave 4 example pictures using AI to generate in that style with 4 standard prompts. One of the prompts was Henry Cavill.
 

FallingDown90

Member
Aug 24, 2018
123
38
The main software is Stable Diffusion. There are mainly 2 user interfaces that is the most established. Automatic 1111 and ComfyUi but there are also others such as Fooocus (a simplifyed UI).
Regardless which user interface you choose, the software is still Stable Diffusion.
There are different model generations, just like there are windows 7, 10 and 11 etc. The ones that are used the most are SD1.5 and SDXL. I would recommend either going with Automatic1111 or Fooocus to get started, when you have become proficient with it, then try a less beginner friendly user interface such as ComfyUi (node based UI) if you wish.

How to install Stable Diffusion Automatic1111:
How to get started Using SD A1111:

Stable Diffusion Fooocus tutorial:

*Edit
Something to be aware of.
Fooocus is primarily using SDXL.
It is supposedly lighter on the vram requirement compared to A1111 though.
I don't know if you can download SD1.5 models and use with it.
Thank you for such clear and helpful answers.
I already used stable diffusion and got used to it, but I was hoping to be able to have real-time results similar to those I saw on leonardo.ai (I know they are different anyway but it would be nice if stable diffusion could use img to img to draw>generate >edit>generate quickly but I think that's too much to ask)
I'm afraid the node interface is too complex for me, I could be wrong because I avoided it with any software....
 
  • Like
Reactions: Mr-Fox

hkennereth

Member
Mar 3, 2019
237
775
Somewhere in here was linked a page that showed examples of all kinds of art styles by artist. I can't for the life of me find it anymore either here or in my history. Anyone got any suggestions that they think might be the one I'm thinking of, or a substitute for it?

Page had a blue background and broke out all kinds of styles by artist, then gave 4 example pictures using AI to generate in that style with 4 standard prompts. One of the prompts was Henry Cavill.
No Henry Cavill here, but maybe this one?
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Somewhere in here was linked a page that showed examples of all kinds of art styles by artist. I can't for the life of me find it anymore either here or in my history. Anyone got any suggestions that they think might be the one I'm thinking of, or a substitute for it?

Page had a blue background and broke out all kinds of styles by artist, then gave 4 example pictures using AI to generate in that style with 4 standard prompts. One of the prompts was Henry Cavill.
It could have been a prompt guide from stable-diffusion- art or similar.


Other possibility, it could be the awesome ebook by Prompt Geek, I have made posts about on several occasions.


A third possibility would be a guide about the styles functions in Automatic 1111, (I didn't find a good example fast).

Also check out the wiki on the github page for Stable Diffusion Automatic 1111.
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
Thank you for such clear and helpful answers.
I already used stable diffusion and got used to it, but I was hoping to be able to have real-time results similar to those I saw on leonardo.ai (I know they are different anyway but it would be nice if stable diffusion could use img to img to draw>generate >edit>generate quickly but I think that's too much to ask)
I'm afraid the node interface is too complex for me, I could be wrong because I avoided it with any software....
This information would have been useful to have when I made the first post. No worries though.
The solution is to use either LCM (Latent Consistency Model) with SD1.5 or Turbo with SDXL. They are both essentially the same at least conceptually in the sense that these models gives high quality with much less steps and a much lower cfg scale. I have made several posts about this lately.
In order to find the models on civitai, simply use the filter for LCM models if you want SD1.5 or search manually.
In order to find SDXL Turbo models, either specify this in the filter or again search manually.
I made a post about this just the other day.
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-12886634

*Try out Fooocus with a SDXL Turbo model. It's a one click install and if you don't like it you can simply remove it after. You likely need to download a turbo model from civitai.
 
Last edited:

Sharinel

Active Member
Dec 23, 2018
598
2,511
Somewhere in here was linked a page that showed examples of all kinds of art styles by artist. I can't for the life of me find it anymore either here or in my history. Anyone got any suggestions that they think might be the one I'm thinking of, or a substitute for it?

Page had a blue background and broke out all kinds of styles by artist, then gave 4 example pictures using AI to generate in that style with 4 standard prompts. One of the prompts was Henry Cavill.
This one I think

 

FallingDown90

Member
Aug 24, 2018
123
38
This information would have been useful to have when I made the first post. No worries though.
The solution is to use either LCM (Latent Consistency Model) with SD1.5 or Turbo with SDXL. They are both essentially the same at least conceptually in the sense that these models gives high quality with much less steps and a much lower cfg scale. I have made several posts about this lately.
In order to find the models on civitai, simply use the filter for LCM models if you want SD1.5 or search manually.
In order to find SDXL Turbo models, either specify this in the filter or again search manually.
I made a post about this just the other day.
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/post-12886634

*Try out Fooocus with a SDXL Turbo model. It's a one click install and if you don't like it you can simply remove it after. You likely need to download a turbo model from civitai.
yes, I'm sorry but believe me, everything you wrote to me was still useful. I studied all the options you suggested and watched your tutorials and I found comfyUI + Local Live a good possibility (for the nodes I saw that they load workflows to order in a more intuitive way)

Now I will evaluate this last piece of advice of yours, I hope I have translated and understood it well XD

Thank you so much for all the help you always give to everyone, you have already helped me with Lora and learning in the past
 

Mr-Fox

Well-Known Member
Jan 24, 2020
1,401
3,802
yes, I'm sorry but believe me, everything you wrote to me was still useful. I studied all the options you suggested and watched your tutorials and I found comfyUI + Local Live a good possibility (for the nodes I saw that they load workflows to order in a more intuitive way)

Now I will evaluate this last piece of advice of yours, I hope I have translated and understood it well XD

Thank you so much for all the help you always give to everyone, you have already helped me with Lora and learning in the past
No worries. You are so very welcome. :)(y)
Don't hesitate to ask for clarifications if you need.
 

DD3DD

Active Member
Apr 23, 2019
858
1,628
I am trying to explain, that the fact that it only happens when SD tries to finalize/save an image that has been upscaled with anything than a tiled decoder node, is giving away a hint about what might be the issue and what might not be the issue, and overheating is definitely not in the pool of possibilities, as it doesn't fit into the pattern.
If you haven't already, maybe just time to do some ole-skool Stress testing with , and . And the other shutdown suspect that gets overlooked, test the PSU with .

P.S. Do update who/what the culprit was?

Edit: I see you suspect the PSU too :geek::giggle:
 
Last edited:
  • Like
Reactions: Fuchsschweif

devilkkw

Member
Mar 17, 2021
324
1,096
Somewhere in here was linked a page that showed examples of all kinds of art styles by artist. I can't for the life of me find it anymore either here or in my history. Anyone got any suggestions that they think might be the one I'm thinking of, or a substitute for it?

Page had a blue background and broke out all kinds of styles by artist, then gave 4 example pictures using AI to generate in that style with 4 standard prompts. One of the prompts was Henry Cavill.
Check and , a good reference for artist and style.
 

Fuchsschweif

Well-Known Member
Sep 24, 2019
1,145
1,957
When I create a picture with a character in the center, but I want some headroom above the character to be not occupied of it, because I want to have space for a title/text, how can I tell SD to leave that room specifically? I have a great seed but the character is almost filling the entire top.
 

hkennereth

Member
Mar 3, 2019
237
775
When I create a picture with a character in the center, but I want some headroom above the character to be not occupied of it, because I want to have space for a title/text, how can I tell SD to leave that room specifically? I have a great seed but the character is almost filling the entire top.
Two methods that work:

1. ControlNet: create or find an image that has the composition you want, and use that as an input
2. Trial-and-error: make many images until you get the right one.

I won't deny that I do the latter a lot more than the former.
 
  • Like
Reactions: Fuchsschweif