[Stable Diffusion] Prompt Sharing and Learning Thread

[felldude]

That sounds very interesting. Do you have any links to what you are talking about so I can do some research?

I think the trained set is behind the devolper wall but the

You must be registered to see the links

So it was Nvidia, Meta, and Amazon, looks like the Nvidia team was the main contributor for it.

I've never done the work to get dali, nvcc and all the other programs needed to use one of Nvidia's sets, they have some of the best in the world though...like the real time set that turns

You must be registered to see the links

The way their models work the CPU and GPU are used and coding is done on the fly, maybe some extremely talented person could convert their pickles to something usable in SD

I did find where they put out Stylegan2 in

You must be registered to see the links

What I find most interesting in that article is they claim Full FP32 training with only using 8.3GB of VRAM using 1024x1024 images....I dont have 48GB of System ram but still (Or a telsa V-100)

You must be registered to see the links

 

[namhoang909]

View attachment 2733142

Language Models

Want AI to write you porn stories? While being completely uncensored, self-guided, using desktop install, and running on a potatoo PC?

Here is the quickest no-frustration guide that will have you running right away:

1. You must be registered to see the links
   - go with "One-click installer" for your OS. Save the zip to the like of: "D:\LMs\oobabooga" and then run the install batch file. Once the install completes you will have an equivalent of A1111 webui installed in the folder from which the batch file was run.
2. You must be registered to see the links
   . Grab the "guanaco-7B.ggmlv3.q4_1.bin" file, that's the actual model, and save it inside "..\oobabooga\text-generation-webui\models".

That's all you need.

Housekeeping regarding settings:

1. Go to tab "Parameters", set "max_new_tokens" to the very max. As of today, the fresh install comes with max of 2000.
2. Go to tab "Chat settings", then subtab "Instruction Template", choose "guanaco non-chat".
3. Go to tab "Text generation", input field, and then: "Write a script for the red riding hood porn film." and watch the magic happen.

PS Also, a relevant video driving the point of why the tokens are important and the race for increasing the token counts:

You must be registered to see the links

PS2 Here is how the LM described a red riding hood outfit for a porn movie that resulted in the picture below:

Spoiler: LM Output

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

View attachment 2733228

Spoiler: More

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

So you use a text to text AI to create the prompt and use that to generate picture in text to image AI, is that right?

 

[Mr-Fox]

Floating Sorceress.. In colab with the eminent Synalon . Don't look too close she might bite..

 

[hanpengyi]

Can anyone share what is the latest cutting-edge NSFW model? I started playing around with stable diffusion today, but it takes years to generate an image (Apple M1 chip), and the online version is not free and doesn't allow NSFW content. I'm especially interested in models that are good at generating fake celebrity pictures, so preferably models with large tranining data on celebrities or have track record of producing good celeb fakes. Thanks.

 

[Mr-Fox]

Most SD1.5 models on civitai does NSFW content. I would recommend Clarity, because it is very consistent. There might be other checkpoint models that produce better images but very few are as consistent as Clarity. For the best result creating celeb fakes either use a Lora for this specific celeb or use the roop extension. With it you are using an image as reference and it is included in the last steps in the generative process. Thank you Synalon for telling me about this extension.

You must be registered to see the links

You must be registered to see the links

examples:

 

[felldude]

Well I did get a setup to train a lora on XLSD, loading latents from disk and using 640,640 instead of my usual 768,768....

10 hours for 1500 steps is about 100 times longer then I am used too though....Im guessing that would be 100 days to train a checkpoint on my machine

Even dropping to 256,256 will take 4 hours

 

[Sepheyer]

Trying to see what will give me Honey Select 2-like backgrounds, and stumbling upon some various pleasant looking pieces:

Spoiler: More

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

 

[Mr-Fox]

Well I did get a setup to train a lora on XLSD, loading latents from disk and using 640,640 instead of my usual 768,768....

10 hours for 1500 steps is about 100 times longer then I am used too though....Im guessing that would be 100 days to train a checkpoint on my machine

Even dropping to 256,256 will take 4 hours
View attachment 2816717

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

 

[felldude]

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

If my math is right you where training around 4.5 sec/it at 768,768 and I am at 30 sec/it at 640,640 and 7-10 seconds per it for 256,256 (It is stretch for me to be training antyhing on a 8GB 3050 at all though)

It is up and available for

You must be registered to see the links

 

[Mr-Fox]

If my math is right you where training around 4.5 sec/it at 768,768 and I am at 30 sec/it at 640,640 and 7-10 seconds per it for 256,256 (It is stretch for me to be training antyhing on a 8GB 3050 at all though)

It is up and available for

You must be registered to see the links

View attachment 2817177

I have only an old pascal card. GTX 1070 Gigabyte Overclocked 8GB Vram. So you should be able to at least match my numbers with a 3050. I'm just guessing, but it would be very strange if a card 2 generations younger isn't better..

 

[felldude]

I have only an old pascal card. GTX 1070 Gigabyte Overclocked 8GB Vram. So you should be able to at least match my numbers with a 3050. I'm just guessing, but it would be very strange if a card 2 generations younger isn't better..

Training on SD 1.5 at 768,768 I average 1 IT/Per second but XL is FP32 the math is exponential

Settings also play a role xformers, buckets etc

 

[Sharinel]

If my math is right you where training around 4.5 sec/it at 768,768 and I am at 30 sec/it at 640,640 and 7-10 seconds per it for 256,256 (It is stretch for me to be training antyhing on a 8GB 3050 at all though)

It is up and available for

You must be registered to see the links

View attachment 2817177

Did I have it wrong? I haven't touched Loras, especially for XL as I was told the base model was 1024x1024. Wouldn't that mean the Loras would have to be the same?

 

[felldude]

Did I have it wrong? I haven't touched Loras, especially for XL as I was told the base model was 1024x1024. Wouldn't that mean the Loras would have to be the same?

It is 1024x1024 but the short answer is no you don't have to train it with the same size image, the image is converted to a latent. For XL most people will probably train with

You must be registered to see the links

 

[Sepheyer]

SDXL in ComfyUI

I am late to SDXL party because procrastination. If anyone reading this thread haven't looked into SDXL, here is a great started video, just watch the first 8 minutes to see what's up. To sum up: SDXL gets text better, it gets better dynamic ranges, better complex composition. And that's merely out of the box. Anyways, here is the video:

You must be registered to see the links

And here ComfyUI example page on how to set up the workflow:

You must be registered to see the links

I haven't tested the workflow yet, prolly a task for this week.

 

[me3]

use other optimizers so you can run a lot less steps. dapt/prodigy works very well and you'll have a much higher rate and save alot of time

 

[me3]

.. To sum up: SDXL gets text better, it gets better dynamic ranges, better complex composition. ...

From the SDXL page:

Limitations

• The model does not achieve perfect photorealism
• The model cannot render legible text
• The model struggles with more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”

somethings seems to no quite add up here...

 

[felldude]

use other optimizers so you can run a lot less steps. dapt/prodigy works very well and you'll have a much higher rate and save alot of time

Adaptive has failed me 4 times with and without xformers, with and without buckets at the recommended learning rate.

I'd try reading this

You must be registered to see the links

, but my gpu is currently and 99% I'd rather my brain not be at the same level.

 

[me3]

Adaptive has failed me 4 times with and without xformers, with and without buckets at the recommended learning rate.

I'd try reading this

You must be registered to see the links

, but my gpu is currently and 99% I'd rather my brain not be at the same level.

with all the hours of testing/training i've run the only "constant" things i've found is that i REALLY wish i had a better computer to run it on and that all the settings/"must use recommendations of YT etc" are basically pointless. Reason being that what works for one or two datasets doesn't work at all for a third. Thinking i'd figured out a setup that worked, as it'd been successfull on 2 datatsets, i ran it on 2 others, one apparently was a car the other went from a blue eyed blonde to a 55yo black woman...

Despite some claims the ia3 (stating that it's mainly for style) using prodigy, works surprisingly well for people, just with simple default values, even without captions and simply feeding it images.

 

[felldude]

From the SDXL page:

somethings seems to no quite add up here...

with all the hours of testing/training i've run the only "constant" things i've found is that i REALLY wish i had a better computer to run it on and that all the settings/"must use recommendations of YT etc" are basically pointless. Reason being that what works for one or two datasets doesn't work at all for a third. Thinking i'd figured out a setup that worked, as it'd been successfull on 2 datatsets, i ran it on 2 others, one apparently was a car the other went from a blue eyed blonde to a 55yo black woman...

Despite some claims the prodify ia3 (stating that it's mainly for style), works surprisingly well for people, just with simple default values, even without captions and simply feeding it images.

One of the best LORA's I have used was listed as an all ages lora and trained on (based on the captions in the lora) nothing but hentai images....I won't call it out but I looked at the settings it used and what I noticed was weighted captions on every image

 

[me3]

If you got the settings to share, i've got some very fun and problematic sets to try it on.
Great fun to see what i can turn some poor ppl into this time, with any luck it might be themselves