Prompt settings to evaluate the candidates and styles, measuring the candidates capable of achieving the a "biomechanical being" and find a "gradient" between the most technological to the most human and another gradient of "features" from each artstyle

Donato Giancola - While "benchmarking" the artists, I'm "minning" for content and the way they "render" the final output, also trying to perceive "patterns" and "artifacting" !
Low CFG scale to find artists that "stand out in the crowd" !!!

Juan Gimenez

Terese Nielsen

Ralph Horsley

Jon Foster

Gerald Brom

Bastien Lecouffe-Deharme

Wayne Barlowe

Hans Ruedi Giger

Luis Royo

Pascal Blanche

Stanley Artgerm Lau

Jesper Ejsing

Bayard Wu

Ed Binkley

Willian Murai

Ross Tran

Chris Rallis

Peter Mohrbacher

Stephan Martiniere

Diego Gisbert Llorens

Diego Jim Burns

Clint Cearley

Philippe Druillet

Jakub Rebelka

Josan Gonzalez

Jean Giraud Moebius

Laurie Greasley

Tomer Hanuka

Olivier Bonhomme

Geof Darrow

Kilian Eng

Alphonse Mucha

Yoshitaka Amano

Tsutomu Nihei

Akihiko Yoshida

Ayami Kojima

Masamune Shirow

Cedric Peyravernay

Charles Vess

Dave Allsop

Frank Frazetta

Hideyuki Kikuchi

Jaime Jones

Juan Pablo Roldan

Justin Gerard

Maciej Kuciara

Marc Simonetti

Michael Sutfin

Noriyoshi Ohrai

Otomo Katsuhiro

Rashed AlAkroka

Slawomir Maniak

sparth

Steve Prescott

Svetlin Velinov

Tatsuyuki Tanaka

Thomas Häfner

Volkan Baga

Wayne England

William Stout

WLOP

Zoltan Boros

Warwick Goble

Edmund Dulac

Arthur Rackham

Kay Nielsen

Aaron Horkey

Alan Lee

Albert Robida

Alex Maleev

Anne Bachelier

Apollinary Vasnetsov

Daniel Merriam

Franklin Booth

George B. Bridgman

Ellen Jewett

After "Minning" all that data, I compressed all these parameters into 2 "carousels" that will "spin" over the generation steps, and set the step count to 48 just to allow the carrousel to loop after all possible combinations

Chose 8 artists, 3 will lead the structure while the other 5 with a more realistic artstyle will "render" that structure!

Some artstyles from these 3 are analog to each other to reinforce the lineart and sharp lines, while Alphonse Mucha will create the background and the borders of the portrait!

Modified the prompt, fixed the "caroussels" by removing their list of values and then blended each artist to preview their synergy and if they can trigger "artifacts" or repetition

After retrieving the blended outputs to "debug" the pipeline, I noticed that Alphonse Mucha isn't going to blend so easy, also, the midle region of the graph has a lot of repetitions, I will replace Ralph Horsley to add more variety to the board!

Added Luis Royo and literaly broke the tiling adding more variety

Alphonse Mucha artstyle tends to mostly take control over the canvas no matter which artist you blend in

Even if you try to "increase" the priority with attention tags you might end up with more os His Artstyle than really let the other artists shine

The order of the keywords prevails over attention tags (it's like giving the canvas firstly to the artist on most left and then that artist pass the canvas with an already painted information, forcing the next one to continue from previous artist )

Now, with that prompt order, we can "pass" the composition of the previous artist to Mucha and allow Him to add His Artstyle ( solving the issue with poses, because Mucha tends to outputs always full body ladies with standing poses )

Reduced the size of the caroussels on artstyle and subject to narrow down the combinations and then start working on solving Mucha's weird bias towards standing poses

Solved part of the issue with some negatives to the prompt

Ok, now the subject isn't looking away anymore, just had to increase the strength of the negative prompt and then added nother keyword specific for "3/4 posing"

Modified the prompt creating a "broad" and "abstract" context, increased attention tags for features that should not take space from the canvas, generated a larger amount of images to check the prompt deviation and check if the fixes are working

Narrow down the negatives for "pose", and instead of fighting with Mucha using Attention tags, we play unfair and "phase-in / phase -out" his artstyle over the steps

That's a comparisson of the "blending modes", the most left tends towards Mucha while the most right tends to a perfect blend with less noisy and chaotic artifacts, while maintaining the beautiful Mucha's nouveau frames around the edges

Also, by "switching" the artists on a caroussel you will notice that "whoever gets first on the initial step of denoising, will dictate the strongest artstyle for the rest of the caroussel"

Also another behavior I'm trying to avoid is loosing Mucha's artstyle, no matter the seed, the amount of steps, the CFG scale, He's fading away!

THAT is the BIAS, a render without any artist, fully rendred ONLY with the rest of our prompt, now, we can see where did the "pattern" previously noticed while "benchmarking the artists" came from! Our prompt is already "converging" into something!

To solve that issue, I should "converge" the prompt towards Alphonse Mucha artstyle without using His artist name keywords, thus, I added nouveau artstyle to make the output slightly converge

Not enough yet...

Then, added "His Era", ( yeah... nouveau art and victorian era almost contained each other )

Almost there, the top line is almost looking like a "retro-futuristic elegant luxurious robot"

Added more negative keywords to avoid "unwanted artifcacts" noticed on previous image

Better ... almost there....

Another additional positive keeyword, more tweaks on negative prompt with attention tags and added a new set of negatives to enhance the output

Good, however, too much, the biomechanical aspect lost a lot of "technological features"

Removed the "Luxurious" keywords and added tweaked the negatives even further

Ok, restored the technology and did not remove these weird corrugated pipes nor the "boldness" ... and I can't use "Ugly" because .... biomechanical looks like a gory dead and ugly also removes that look

Tweaked again the negatives, this time with everything that could "trigger" the perceptron to remove these damn corrugated pipes...

Yeah... time to move on ... I'm giving up and let the artists remove the probability of these damn corrugated pipes and a bald head

Maybe reorganizing the core structure of the subject!
Also, I've removed the useless attempt of avoiding these ugly corrugated pipes

!!! FINALLY !!!!
I know, it's looking "ugly" and it's not "pleasant" to the eyes, however, the "content", the features and the color palette are finally blended !!!

Created the caroussels and added an importance tag on specific keywords, trying to allow them to kick in even after many steps of delay

Finally some colorful Vintage Retro-futuristic biomechanical portraits!

Also, if I try to add the third artist right after the second, Alphonse Mucha will lose/skip 2 frames/steps of denoising, thus, for that reason, I added Alphonse Mucha between these 2!

Using an attention tag by the end of the caroussel will end up with a pipeline that "converges" all artists into the same output (no matter which artist, and I don't even know why)

The output

However, increasing the attention tag over Alphonse Mucha, we might get a more noticeable impact of every artist

The output

Adding Attention tags to Mucha twice in the caroussel will not produce something coherent and everything is going to look "too glitchy"

The output

Also, I could attach more slots within the carroussel using the same pattern : "Alphonse mucha |New Artist name"

The Output

And keep pushing more and more artists into the caroussel

The Output

Creating a large caroussel mixing all artists might probably end up with something very detailed and chaotic, while using only a caroussel of 3 diffferent artists produces an output controlled, coherent and pronounced artstyle

Then, if i maintain Alphonse Mucha as the "art director" while I swap his "lead artist" in charge of creating the structure, and the other artist in charge of the final render, I can achieve with this "pipeline" different rendering / content options

The outputs

Then, it's time to "unpack" every artist "hidden feature".... since stable diffusion trained mainly on 512x512 images, rendering 384x576 will only generate images with the "minimal set of features", thus, upscaling to 576x864 ...

Then... pushing the resolution to 640x960 not only increase the amount of detail and adds new features, but start "tiling" vertically the features and even change the final rendering style

Since higher scales brings you more details and a different rendering, the inverse also achieves less detailed images, and, depending on the seed and denoiser, without enough steps, they will hardly match the prompt!

2 outputs with noticeable less details

with low upscaling, a huge denoise strength outputs "cleaner outputs"

The Output

With higher resolution, a lower denoising strength seems "healthy"

The output

Another last sample with all Blends

After saving the "bare bones" of our pipeline, it's just a matter of customizing it to achieve different results

Awesome.... and .... how it's going to look like on other custom checkpoints of stable diffusion ?

Then... time to test as many checkpoints as possible to test these "virtual artists" artstyle over the prompt