Another Damn Art Model (ADAM)

Questions/Feedback/Updates?

Visit my thread on the Unstable Diffusion Discord

Like Fantasy & Sci-fi Art? Check Out My New Model : Doomer Boomer

This merge is made up of different epochs and iterations of sample models I created to test a much larger fine tuned model I'm working on (Boomer art model).

Artist tags from the 1980s and 1990s Heavy Metal Magazine era can really improve the overall composition and style of the output generation.

List of Artists to Try :

Boris Vallejo
Julie Bell
Boris Vallejo and Julie Bell
Frank Frazetta
Luis Royo
Brom
Richard Corben
Larry Elmore
Olivia de Berardinis
Scott Hampton
Moebius
Milo Manara
Vince Gimenez (or just 'Gimenez')
Chris Foss
Philippe Druillet
Richard Cohen
Doug Chiang (Great for sci-fi landscapes)
H.R. Giger
Hajime Sorayama (or just 'Sorayama')
+ more

Some artist tags will be more influential than others, this shouldn't be as much as a problem when the full model is released. Experiment with other artists not listed above, even some modern artists are known by the model. For example 'background by Jessica Rossier' can create some very beautiful sci-fi landscapes. If the artist style isn't pushing through enough, look down bellow and read "Neutral Prompt Hack".

Requirements

VAE Required: Download here

Install :

stable-diffusion-webui -> models -> VAE

In Webui :

Under Settings -> Stable Diffusion -> SD VAE select the downloaded VAE and uncheck Ignore selected VAE for stable diffusion checkpoints that have their own .vae.pt next to them

Settings

Clip Skip : 1

Settings -> Sampler parameters -> Clip skip

ETA Noise Seed Delta : 31337

Settings -> Sampler parameters -> Eta noise seed delta

Face Restoration : Disabled (See ADetailer)

IF enabled : Settings -> Face restoration -> CodeFormer -> weight = 0.8

Parameters :

Sampler :
- I used a combination of Euler a, DPM++ SDE Karras for many of the sample images, but this is just personal preference.
Steps : Depends on the sampler, typically 20 - 35 works well.
- If DDIM use higher steps

Hires Fix (Optional) :

Upscaler :
- R-ESRGAN 4x+
  - Denoising strength : 0.3 - 0.35
  - Upscale by : 1.5 - 2 (can go lower but shouldn't go higher than 2)
- Lolipop, 4x-UltraSharp, Remacri (foolhardy)
  - Can be downloaded here
  - Install to stable-diffusion-webui -> models -> ESRGAN
  - Typically Denoise ~= 0.5 - 0.6 (can go higher or lower)
  - Upscale by : 1.5 - 2

Extensions

ADetailer :

Extension for webui that uses ultralytics to 'fix' faces/hands/body after generation.

Download here - https://github.com/Bing-su/adetailer

Read the information in the Github repo linked above for more info and installation guide.

Neutral Prompt :

Download here - https://github.com/ljleb/sd-webui-neutral-prompt

Note : If you look at gen data for the sample images and see AND_PREP or AND_SALT; then the image used this extension

Neutral Prompt Hack : Okay, it's not really a hack. But I use neutral prompt in a way that it wasn't necessarily designed for.

Let's say you want to really force a specific art style. Open the neutral prompt block in webui and expand the "Format" option. This gives you a CFG slider, prompt box, weight slider, and drop down menu. Leave the dropdown set to "Perpendicular". In the prompt type the artist who's style you're trying to force. For example, by Boris Vallejo. Leave the weight slider at 1 for now, you can change that later if needed. Change the CFG slider (in the extension block) to 0.5. Click the button to append these settings to your prompt.

Your main positive prompt should now end with something like : AND_PREP by Boris Vallejo : 1

Click Generate, it may take a little longer but you should see the art style of the Artist being forced.

This shouldn't really work, and I will need to dig deeper into the math to figure out why it does. I'll update this when I do.

Your results may vary, and for this specific model—it's not really needed. But it's there if you want to give it a try.

Feel free to modify the parameters of the neutral prompt extension, nothing is set in stone.