The MadArt archive starts on August 13, 2022, with a string of cartoon-raccoon prompts that produced 728 images in a single day. Mad Bitcoins was using MidJourney v3, the version with the soft watercolor defaults and the famously specific colorblock palette. Three years later, the same prompts run through ChatGPT (GPT-image-1) and Google Gemini produce something almost categorically different.
The clearest single comparison is the iconic prompt — "cartoon racoon samurai" — which Mad Bitcoins ran 64 times across August 2022:
The 2022 MidJourney result is a flat-color print, framed like an old fortune-teller card, with a small symbolic raccoon foregrounding a stylized city of red moons and Japanese placards. The two 2026 results are full-bleed cinematic illustrations: armored samurai-raccoon characters mid-swing over photoreal Tokyo, with legible signage, lacquer-shaded armor, lighting passes. The medium changed from poster to scene.
Three shifts you can see in every triplet
1. Photoreal-by-default. The 2026 models route to a finished commercial-illustration look unless you fight them out of it. MidJourney v3 had a strong painterly default; v5 cleaned that up; the 2026 generation has standardized on "polished product art." The raccoon-on-the-moon comparison makes this concrete:
MJ gave a melancholy stylized illustration. ChatGPT gave a Pixar-poster astronaut raccoon with a Saturn-V rocket; Gemini gave a children's-book raccoon holding a flag that reads "I ♥ THE MOON!". Same nine-word prompt, three completely different intents.
2. Prompt adherence is much stronger — and louder. Where MJ 2022 would interpret "a racoon version of speedracer in a circle for a logo" as a literal logo (clean geometric raccoon-in-circle, three colors), 2026 reads it as a brief for a vintage merchandise badge:
ChatGPT delivers "RACCOON Speed Co." text, a red sports car with the number 7, a turbo gauge, lightning bolts. Gemini delivers "SPEEDRACCOON RACING CLUB." Both interpreted "logo" as "branded sticker" rather than "mark." Prompt adherence has improved on every axis except restraint.
3. Text-in-image now works. Every 2026 image in this pilot contains some legible English: chalkboards, signage, banners, t-shirt prints. The MidJourney v3/v4 originals almost never do — and when MJ did try, the letters dissolved into texture. Compare the breakfast scene:
The MJ raccoon holds a small plate against a beige void. The ChatGPT version is a complete cottage kitchen with a chalkboard reading "Today's Plan: Eat Pancakes / Be Happy / Repeat," labeled FLOUR and MILK containers, and an apron embroidered "Good Morning!". Three years ago that signage would have been illegible noise.
The rest of this pilot — eleven collections, thirty-five prompts — extends those three observations. The next three essays go deeper on each.