Mars College AI Art.

AI Art generated by Marzipan at Mars College [http://mars.college] Images generated from text prompts using http://eden.art (stable diffusion)

[falsecalligraphy]

Fatso Trump in Alice-in-Wonderland drag, down a rabbit hole, promoting a DMT feverdream of obesity, sponsored by hamburgers

[falsecalligraphy]

"With A Stronger Dose":

For these images, I used a prompt that contained a quote from the book I'm reading:

"With a stronger dose, a state of fear and anxiety could be induced where the subject could be compelled to trade their cooperation for a guarantee of return to normalcy." (Acid Dreams: The Complete Social History of LSD: The CIA, The Sixties, and Beyond, by Martin Lee (1992))

I edited the quote to make it more suitable for prompin: "(A state of fear and anxiety, induced by a stronger dose). Compelled cooperation, (a promised guarantee of return to normalcy)."

Some of my secret sauce is to use style qualifiers at the back of the prompt (after the quote). Here is what I used for these: "Lysergic forced perspective, DMT linear perspective, tryptamine tilt-shift, dazzling ayahuasca hallucinatory scene, in a style reminiscent of <concept>."

I made these with Eden (app.eden.art), applying different LORAs ("concepts" in Eden), trained on different sets of images: stills from the movie "Holy Mountain" by Alejandro Jodorowsky, synthetic bio-inspired remixes of the "Mars College" logo, patterns of boot-imprints found in the dust on a dry lakebed (Salton Sea, CA), and the Burning Man man-base effigy.

The series (with more variations) can be viewed here:

Collection "withAStrongerDose" by marzipan | Garden of Eden

39 creations curated by marzipan

[falsecalligraphy]

[falsecalligraphy]

[falsecalligraphy]

[falsecalligraphy]

[falsecalligraphy]

[falsecalligraphy]

[falsecalligraphy]

Two different sets of variations, using "concept-shopping", as part of this week's Sunday Miracles challenge on http://app.eden.art -- trying out different LORA concepts that people have finetuned.

Here I used a photo of my friends carrying a couch that had graffiti on it, and then a photo of me collecting bacteria samples from the exposed sediments of the Salton Sea, yesterday morning (photos used as controlnet guidance images).

[falsecalligraphy]

Here are some images I created based on Love Tester, a Burning Man honorarium piece I created (and built with my besties) in 2015. These are created using a "concept" (aka LoRA) in app.eden.art, from about eight photos of the finished installation. The bluish ones use a weak concept strength as well as a prompt that uses the words "cyberpunk", and the last one uses a high concept strength, and kinda resembles the piece we actually built remarkably well.

Since we actually burned the piece at the end of the BM2015 event, I am really fond of getting to see these images because otherwise the project just lives on in my memory. I really appreciate how AI allows us to resurrect half-life specters of bodies that have physically departed this world, and also to explore diversified "cousins" that vary on a theme. It is super exciting how these tools can help artists and builders ideate and iterate.

[falsecalligraphy]

More AI remixes of the Mars College logo, with diatom/foraminifera prompts

[falsecalligraphy]

More compositions based on the Mars College logo concept, these are very microbiology-inspired, but some of them also resemble the vascular cambium of higher plants (which I did not try prompting, but should!!). The first one was inspired by carboxysomes, subcellular microcompartments which self-assemble into icosahedral protein cages packed with the rubisco protein, which help cyanobacteria fix CO2 more efficiently. (I have done a lot of work on carboxysomes in my science research).

[falsecalligraphy]

Some remixes of the logo for Mars College, using a LORA I trained on it, and various different promp-craft

[falsecalligraphy]

Here are more variants of album covers (of experimental post-rock records I am quite fond of), using the "controlnet" function of app.eden.art ("canny edge" option). These sets explore the higher range of controlnet strengths, preserving more of the original image composition.

I really love the way the Sangam ones exploit the different angles of lines that overlay the original album cover, as well as the different positionings of where it makes my likeness appear in the set from Melodien's "Apeiron". The moiré effects of the set from Akasha System get used in some really cool ways.

In all of these, I use a LORA trained on myself (at 80% concept strength) as well as identical text prompts (something about a cybernetic guru from a futuristic hippy commune trying to avoid distractions from too much technology while meditating about the hive mind).

[falsecalligraphy]

Here I am exploring the "controlnet" function of app.eden.art, where I use different album covers of records I like, as initial images (as the "init_img") which get used to guide the composition. I show decreasing control strengths of the init_img -- notice how the pictures in the right column are more similar to each other because they are less constrained by the album cover.

It is interesting how the album covers that are purely geometric (Snow Palms & MinaeMinae) do not show any of my likeness at the highest controlnet strength (init80).

I love how the multipanel boxes of the Sonic Youth cover (from the album "Sister") begin to dissolve as the controlnet strength gets relaxed.

For all of these, I used the "canny edge" function, as well as a LORA of myself based on six selfies as training images. All used the same text prompt.

[falsecalligraphy]

Using the "blend" function of app.eden.art to combine the runway looks of Michelle Visage and RuPaul (from episodes of Drag Race UK Season 5).

It is interesting how most of the interpreted prompts detect that these are screenshots from Drag Race. They never seem to identify Michelle correctly, although it does interpret her "queer makeup". I think Ru would be happy to discover that several of these creations resemble Vanessa Williams.

[falsecalligraphy]

Here are a few of the iconic and amusing (albeit disturbing) dirty secrets that I illustrated using eden.art for my Dirty Secrets Illustrated zine project, for giving out in association with our art installation this year ("Mini Dirty Secret") at Burning Man 2023. We have been collecting these anonymously since 2017.

An interesting thing that I discovered is that the worksafe-features of the SDXL model not only prevent pornographic image creation but also seem to be unable to depict the actual act of violence or cruelty (I couldn't get the kitten to appear trapped underneath a shoe, I couldn't depict the nine-year-old in the act of striking the toddler nor the guy actually laying his fist on his mother) as well as defecation. I think I got the urination picture using an nsfw checkpoint that someone finetuned (hosted at seaArt), but I had to specifically promp that a puddle was at their feet and that a stream of liquid was pouring down their pants leg.

I'm writing an in-depth blog post about these "AI safety" aspects, which will appear on my substack. I learned so much about prompin' from this project, and I find it fascinating that today's models prevent a lot of vulgar (but not necessarily nsfw) imagery to be created with AI. Many of these images took me dozens of trial-and-error attempts to devise a suitable text prompt.