Stable Diffusion AI Image Generator

It doesn't seem drastic, maybe a 15% increase if I had to guess... I see it loaded in the console and looks like it applies with each image generation and upscale too.

Launching Web UI with arguments: --xformers
Patching transformers to fix kwargs errors.
Dreambooth API layer loaded
LatentDiffusion: Running in eps-prediction mode
DiffusionWrapper has 859.52 M params.
Loading weights [81761151] from D:\StableDiffusion\stable-diffusion-webui\models\Stable-diffusion\model.ckpt
Loading VAE weights from: D:\StableDiffusion\stable-diffusion-webui\models\VAE\vae-ft-ema-560000-ema-pruned.ckpt
Applying xformers cross attention optimization.
Model loaded.
Loaded a total of 2 textual inversion embeddings
 
Last edited:
SD2.1 and 2.1_base has been released.
You can use the same yaml from 2.0, just rename.
 
With all the negative backlash SD2.0 had recieved, I haven't even played with it yet, there's still so much to learn with 1.5, creating custom Dreambooths, Deforum animations as well as InvokeAI with its outpainting tools. I'll probably check out 2.1 pretty soon though.
 
OnlyOneKenobi said:
With all the negative backlash SD2.0 had recieved, I haven't even played with it yet, there's still so much to learn with 1.5, creating custom Dreambooths, Deforum animations as well as InvokeAI with its outpainting tools. I'll probably check out 2.1 pretty soon though.
Click to expand...
The issue with 2.0 is that there was an active effort to censor a large portion of the model because the creators didn't like how it was being used.
That's not what the community expects, nor what they wanted out of these models, hence the sizeable backlash.

2.1 made large steps to address that issue but it's still not 100% thus why people are, even with the almost 100% reintroduction, still sitting on 1.5.

Unfortunately, or fortunately if you're so inclined, like every new technology, porn will always be a driving force of new tech. AI for the masses is no different.
 
Last edited:
Vorastra said:
The issue with 2.0 is that there was an active effort to censor a large portion of the model because the creators didn't like how it was being used.
That's not what the community expects, nor what they wanted out of these models, hence the sizeable backlash.

2.1 made large steps to address that issue but it's still not 100% thus why people are, even with the almost 100% reintroduction, still sitting on 1.5.

Unfortunately, or fortunately if you're so inclined, like every new technology, porn will always be a driving force of new tech. AI for the masses is no different.
Click to expand...
I've checked out 2.1 briefly. It's easy enough to install and get running, etc. but I don't see much improvement in the landscapes that I've generated with it. Either my 1.5 landscapes were already that good, or both my 1.5 and 2.1 attempts are bad, I just can't decide. I haven't even bothered with trying to do portraits or animals using 2.1 yet since its selling point is awesome landscapes.
 
OnlyOneKenobi said:
I've checked out 2.1 briefly. It's easy enough to install and get running, etc. but I don't see much improvement in the landscapes that I've generated with it. Either my 1.5 landscapes were already that good, or both my 1.5 and 2.1 attempts are bad, I just can't decide. I haven't even bothered with trying to do portraits or animals using 2.1 yet since its selling point is awesome landscapes.
Click to expand...
Yeah I haven't noticed any improvements in generation quality, even though people on Reddit post supposed improvements, with 2.1.
2.1 is slightly faster but not by much, around 8%.
 
Been playing with a new model from a fellow on Reddit.
Seems to be really good for portraits.

Analog Diffusion.

Euler A 20 steps, about twice as slow than normal SD1.5 but the results are pretty good.

EDIT: seems it was just the first run that was slow. Subsequent runs are just as fast as normal SD1.5.

Here's some of the results from literally the first run.

1670755681679.png 1670755748750.png
1670755767807.png 1670755775379.png
1670755781469.png 1670755800134.png
 
Last edited:
So I've been dusting off my dormant powershell and python scripting abilities to make a "face blender" and "style builder" tool for generating portraits of consistent custom characters in SD.
A lot of criticism people on Reddit seem to have for the technique of blending faces to make custom characters is that if you only blend two or three well known actors, is that your character is still somewhat recognisable as either or all of those actors instead of being a truly original face.
So my experiments have led me to believe that if you blend three or more faces, as well as set a custom defined "style" consisting of age, gender, hair colour, hair style, expression, clothing, and physique, it's pretty dang difficult to look at these faces and identify the actors who've been blended together. To take it even further my scripts can blend many more faces than just three or more in different ratios and then swop genders, change their age and physique (I haven't even tested more than 6-7 on average) and then you end up with truly original characters that can be repeated in multiple generations too.
Sure it's not 100% consistent and sometimes there is too much variance from one pic to the next, but I'd say one can at least get 70% where the character will look like the same person, even if their age, clothes, and hair style maybe changes slightly from picture to picture. If more consistency is needed, then pictures from these batches can be cherry picked and used to create a custom Dreambooth model for that character.

01 - the young woman
BlendedChar01.jpg

02 - the middle aged serial killer
BlendedChar02.jpg

03 - the fifty something year old politician... sometimes he has a goatee, but usually just a moustache.
BlendedChar03.jpg

04 - somebody's cute kid... pic 3 doesn't really look like her at all, but like I said, it's not 100% consistent all of the time.
BlendedChar04.jpg
 
So, an interesting thing about this Analog Diffusion model.

The guy says he didn't use any nude images whatsoever in training it, but if you don't put the "nude" negative prompt, let's just say you get...interesting results.

ngl, it does a good job. I think the guy on Reddit might be lying about how he trained it. :unsure:
 
Yeah, the Analog model is pretty cool. I helped it along with some img2img upscaling in SD and Photoshop, and then I applied an additional film grain over it because I will always tinker with photos before I'm completely happy with them, but yeah, the base image was pretty ok to begin with. I haven't had any NSFW results (yet)

Analog.jpg
 
Has anybody managed to get SD/webui working with AMD(RX580) on Windows 10? I've been battling to get it going - even tried booting Ubuntu from a USB but not coming right so far.
 
InvisibleJim said:
Thanks, I tried this. It was going OK up until downloading the model but I can't see the actual model itself in the onnx branch of Huggingface. Very frustrating.

I also tried with SHARK but it throws an exception and fails when I run the script. :-(
Click to expand...
As an alternative to a local install you could run it from Google Colab, it usually works, even with free membership. It's just a few more hoops to jump through every time you start up Stable Diffusion, but if all else fails, it's a worthwhile alternative.

There's a lot of notebooks set up here that seem to be configured for the most popular SD models.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X