AI Images with Interesting Aesthetics

[CycFL]

Has anyone ever noticed how the vast majority of AI images posted anywhere have almost the exact same aesthetic, despite the fact that they are capable of producing a wide variety of them? Why is this the case?

I think image models are interesting because they contain a representation of the human visual world from a fundamentally nonhuman perspective (which is why I don't like finetuned models as much; they restrict it to a small subset of such). It makes you wonder what patterns hide within...

Anyway, post AI images that don't have that aesthetic. The more unique the better.
I'll start:

 

[Some_porcupine]

Has anyone ever noticed how the vast majority of AI images posted anywhere have almost the exact same aesthetic, despite the fact that they are capable of producing a wide variety of them? Why is this the case?

I think image models are interesting because they contain a representation of the human visual world from a fundamentally nonhuman perspective (which is why I don't like finetuned models as much; they restrict it to a small subset of such). It makes you wonder what patterns hide within...

Anyway, post AI images that don't have that aesthetic. The more unique the better.
I'll start:
View attachment 69387View attachment 69388View attachment 69389View attachment 69390

maybe they are biased and just really like (those AIs) weirdcore, oddcore and "fleshstarr" lol

 

[MySpace Tom]

It's probably because everyone is accessing instances of the same model and using queries that are very similar to one another


Recently, Midjourney introduced the Zoom Out feature. It is much jankier than if you create an entirely new image from a prompt. I was having it generate weird cartoon characters when I tried it, but when I had it zoom out, it straight up put a bunch of Naruto copies in one of the images

This is the input image before zoom:


And these are the four variations it returned after I applied zoom:

 

[rémineur]

To me, images that are generated by Dall-E are often more interesting because they are less perfect than the ones made by Midjourney. And because of this imperfection, some images have a pretty unique aesthetic. Here some exemples :

 

[alCannium27]

Speaking of AI generated images, base Stable Diffusion models are pretty bad at generating none-cripples, so most "professional" AI art enthusiasts uses one of several negative embeddings to avoid stuffs like fused fingers and such. Also at least as far as SD goes, I think most "pros" uses shared wildcards, which are collections of terms that form certain types of images, mostly either anime or photorealistic, and going by what most anime waifu posts on Civitai, everybody's using Danbooru tags so anime images would all look similar. Besides, it's anime, they all look the same, every single one of them, NO EXCEPTIONS!

I think in this regard people are generating images with a goal in mind, either with existing templates (for example, I use only photo realistic SD models and has the phrase "HD digital photograp of {prompt}, high quality, intricate detail, professional lighting" as the base prompt for everything, as a result, every image generated looks like a studio photo shoot) or specialised addons (like LoRAs of an artist or a person, which will result in a style very similar to the training data).

I had generate stuffs like Lenin and Stalin wearing fallout style power armor standing in front factories and such, those always look like old colorized photos from the 30s (because the training data certainly are). Such is the cost of using LoRAs I suppose.

EDIT: Also my portrait/avatar/whatever you kids call it these days is one of the first AI image I generated with SD1.5, trying to get it do something like Metropolis

 

[CycFL]

more

 

[CycFL]

Was playing with SDXL recently. It's definitely harder to get stranger aesthetics, but certainly not impossible.

 

[alCannium27]

Damn, those Roman numerals on that clock was almost half correct. So sad, maybe only GAN can do text correctly.

 

[Some_porcupine]

ai starts to draw from ai prompts now
degeneracy

 

[alCannium27]

ai starts to draw from ai prompts now
degeneracy

Small if true.
Glory to the machine. AI should have taken over by now if the world is just.

 

[MySpace Tom]

ai starts to draw from ai prompts now
degeneracy

The Singularity has arrived, and it's entirely underwhelming

 

[Elagabalus]

Some early Midjourney had a nice aesthetic. The latest version can make very photorealistic so with some tweaking there's some decent "art" it does.

 

[alCannium27]

Here, a LoRA I used to get a few images in SD:

Spoiler: We Got the WTC At Home

Positive Prompt: High-resolution Photorealistic Image of View of the new york twin towers from across the Hudson river <lora:1947-Paris-SD1.5-5:1> 1947-Paris-France, bokeh, LDSR, Depth of Field, OverallDetail
Negative prompt: easynegative, verybadimagenegative

Spoiler: Typical Parisian Economy

Positive Prompt: High-resolution Photorealistic Image of <lora:1947-Paris-SD1.5-5:0.8> 1947-Paris-France, photo of a faded movie poster featuring a starlet's face in profile against a gloomy dark backdrop, outside on an empty rainy street. gloomy sky, clouds, bokeh, LDSR, Depth of Field, OverallDetail
Negative prompt: easynegative, verybadimagenegative

Actually used a different prompt to get the street and slapped the poster on the wall, than I2I for the results. SD1.5 is about half as smart as SDXL and SDXL ain't exactly on passing grade.

Spoiler: I NEVER ASKED FOR THE DAMN TOWER

Positive Prompt:High-resolution Photorealistic Image of <lora:1947-Paris-SD1.5-5:0.8> 1947-Paris-France, photo of a simple tumbler on a see-through glass table by a large window, clear blue sky outside, bokeh, LDSR, Depth of Field, OverallDetail
Negative prompt: easynegative, verybadimagenegative

Spoiler: NORMALITY