Comparing AI-generated images two years apart — 2022 vs. 2024.

I have been using various AI Image Generators occasionally since the Summer of 2022 — mostly Midjourney, Stable Diffusion, DALL-E, and Leonardo AI, though I did not use them as often as most enthusiasts. The benefit I enjoyed was mainly in generating the illustrations and background images for my lecture and presentation slides. It is lame, I know. But it provides me an opportunity to keep observing the development of such tools.

AI toolkits are evolving like rockets. The quality of AI-generated images has gone from OK to jaw-dropping in only 2 years. How far did we go exactly? A side-by-side comparison will tell.

I dug out some images I generated in Midjourney in 2022–2023 and used the same prompts to regenerate them in the latest version (v6) of Midjourney. You be the judge. BTW, as you can see, I am not a big fan of long and sophisticated prompts. That’s just me.

PROMPT: an animal with mechanical and steampunk shape in the forest under rays of light.

In both v2 and v6, they are all cool, as long as it is steampunk. I actually like the vague and dreamy silhouette in v2. The quality and details in v6 are stunning.

PROMPT: an apartment of 2 bedrooms, large kitchen, living room, with balcony, in French style, facing South, in architectural plan view.

Midjourney v2 knew nothing about modern architectural drawings. What it knew resembled the collage of lines and shades taken from drawings it learned from, nothing else. Architecture? Not even close. In v6, it can do architecture drawing quite well but still lacks the basic understanding of a reasonable spatial layout and requirements. For example, where is the toilet? And the openings are either missing or misplaced.

PROMPT: high rise building, thin towers, city scene, colorful rooftop, in Kandinsky style.

Kandinsky would have killed me. But I like both of them. I think kids love them, too.

PROMPT: grand piano keyboard, closeup view, under forest shades, volumetric light, realistic, depth-of-field.

Midjourney v2 was horrible — quite an impressionism. In v6, the layout of subjects, image quality, and rendering effects is fantastic. BUT aren’t there too many ebony keys? The reflections on ivory keys are too much.

PROMPT: A city, there is the great roller coaster with its steep humps, the carousel with its chain spokes, the Ferris wheel of spinning cages, the death-ride with crouching motorcyclists, the big top with the clump of trapezes hanging in the middle.

The details in v6 is crazy… WOW. The composition is a bit messy, though.

PROMPT: in the urban jungle, a gigantic mother tree with sophisticated branches connecting everywhere, dangling fruits like fireflies, ultra realistic, volumetric lights, 4K. Trending on ArtStation. Sci-fi style.

I have always been curious about how Midjourney expresses “shimmering light.” Quite satisfied with v6.

PROMPT: a cyborg plant with circuit and networking capability, close-up, realistic, hi-res.

Cyborg vegetation is a subject that I have been investigating in recent years, from imagination to implementation. In v2 the composition was simple, way too simple. In v6 the composition is sophisticated and a bit too overfull. The contrast of the image can be further improved.

PROMPT: a beautiful artificial plant with flower petals made of airbags and sticks with embedded electronic circuits. In a cyborg nature.

I guess Midjourney v2 din’t know how to generate airbags, but it is resolved in v6. The translucent airbag petals are beautiful!

PROMPT: a poster design titled “Maker Faire Taipei 2022”, with a joyful and rich playful feeling, children’s picture book style, and red robots helping kids work on creative projects.

You can see how frustrating it was if you wanted to generate a poster design in v2. The inability to generate non-English text is still a shortcoming in v6.

PROMPT: a cute rabbit eating a moon cake.

The definition of “cute” in v2 totally went wrong. The rabbits were like characters in The Addams Family.

PROMPT: An art craft by Melpomeni Chatzipanagiotou, a brown bear and a white rabbit sitting together in a landscape, a forest on the far side.

I still remember the hyperactive reactions, if not ecstasy, of the world toward AI Generative Images in the Summer of 2022. Now, when I see the vintage images done by v2 in 2022, I can’t help but wonder why.

PROMPT: a large shiny yellow sphere with a large X crack on its upper surface, blur background.

The v3 images were like…hmm…it only caught the theme. On the other hand, in the v6 images I didn’t expect such an exquisite quality of the overall image and the crack. AI-assisted rendering is definitely an exciting future.

PROMPT: a photograph of a convenience store on a rainy night, shimmering light, overexposure, glitch style.

PROMPT: an Asian convenience store house in the countryside, by the street corner, fine details, jaggy bold strokes, white background, in color pencil sketch.

I thought colored pencil meant dry colored pencil, but the first image by v6 looks like water-soluble colored pencil or watercolor.

The following images were generated in v5.1 or newer versions, which showed muuuuuuch better quality than v2 and v3. The main difference between v5.1 and v6 is the accuracy and integrity of the interpretations of the prompt.

PROMPT: a complex morphing machine inspired by Amoeba. White background.

PROMPT: a white tower of diagrid structure made of plastic air pipes and pneumatic joints. black background.

PROMPT: a stage lighting scene using the Fall color palettes. Dark background.

IMHO, Midjourney v6 is qualified as an image generator (or interpreter) for professional applications in most cases, no matter how bad your prompts are.