Cool Pictures from the Mind of a Machine: AI Generated Pictures

[Thorgalaeg]

There are a lot of different models and Lora based on SDXL now, and except maybe upscaling using ControlNet, it should be able to do anything that SD1.5 could. But prompting it is different, you can't just re-use 1.5 prompts with SDXL.

There is a new controlnet tile upscale model for SDXL suitable for realistic images (until now there was one for manga images only). Havent tried it myself though. You can find it here: https://huggingface.co/bdsqlsz/qinglong_controlnet-lllite/tree/main
The last one.

 

[EvaDK]

In the start it was extra fun to really dive into AI generated photo realistic images and find mistakes, like a staircase leading up to a wall where a door would have been, or a celebrity with 6 fingers on one of his hands instead of 5.

 

[Thorgalaeg]

In the start it was extra fun to really dive into AI generated photo realistic images and find mistakes, like a staircase leading up to a wall where a door would have been, or a celebrity with 6 fingers on one of his hands instead of 5.

Still happens a lot. Current image generation AI has not a real understanding of human anatomy (or of anything for that way) it basically converts pure noise in spots and finally in images in some kind of super pareidolia, using the info it has from the billions of captioned images used for training it. Fine-tuned models can improve that but wont solve it completely. Even the awesome Sora (text to video) have issues with real world physics:

The guys at OpenAI say it is only the beginning though and things can improve quickly.

 

[Valka D'Ur]

Stormy Saturday night here. So some random relaxing landscape pics are in order:

View attachment 685759View attachment 685760View attachment 685761View attachment 685762View attachment 685763View attachment 685764View attachment 685765View attachment 685766

The impressive part is all these images were generated in a few seconds by my 3090ti using a new kind of checkpoint (SDXL Turbo) which only needs as few as 3 steps to fully render an image when a minimun of 15-20 steps have been needed till now. In fact now it is possible to generate and retouch AI images in real time.

The 6th one has a broken column.

 

[Bonyduck Campersang]

From here.

I never cease to be amazed at how the AI manages to generate images of text, especially when it comes to fancy typefaces

 

[Birdjaguar]

These are concert posters from the late 1960s. Three of hundreds that were made by artists. Not AI created.

 

[Bonyduck Campersang]

These are concert posters from the late 1960s. Three of hundreds that were made by artists. Not AI created.




View attachment 686073 View attachment 686074 View attachment 686075

Yes, these are of course far superior to anything AI can generate. What I was talking about being amazed by is the fact that AI is not a thinking machine, and it's not typing text in a fancy font but generating an image of the text. So what the AI has to do when given a prompt with instructions to generate a piece of text, it has to generate an image of the text, rather than simply typing it out, and then it has to do it in a fancy font, which again is not an embedded font within the AI programme but the AI's 'idea' of what a particular font is, and has to match the shape and style of the text to that font, and retain consistency of style. This is something humans of course are very good at doing, but a machine doing it is a very different thing.

 

[Smellincoffee]

Being perverse with Bing image creator. Prompt was:

Draw one American football player running with the football being chased by another football player; the first football player is in a blue jersey with "RED SOX" written in red; the second football player is wearing a white jersey with blue pinstripes and blue letters that say "NEW YORK".

 

[Thorgalaeg]

Weird randomness:

Spoiler :

 

[Gedemon]

A few pictures from my latest "exoplanets" generations (2560x1440)

Spoiler more... :

 

[Bonyduck Campersang]

Being perverse with Bing image creator. Prompt was:

The image doesn't show up for me in your post but I can see the link for it when quoting it

 

[Lexicus]

An AI-generated ‘Balenciaga pope’ fooled us all.

Did it "fool" us though? I never thought it was real, did anyone?

 

[Comrade Ceasefire]

I nudibranchs.

 

[Bonyduck Campersang]

 

[Comrade Ceasefire]

View attachment 686676

Johnny Hendrix guy was a helluva guitarist, even when using only 5 strings.
I like the way you made him look as if he was hazing that Prince twerp who would have needed 6 strings.

 

[Birdjaguar]

Jimmy not Johnny.

 

[Bonyduck Campersang]

Johnny Hendrix guy was a helluva guitarist, even when using only 5 strings.
I like the way you made him look as if he was hazing that Prince twerp who would have needed 6 strings.

I can't tell whether you're joking, but the person in that image is meant to be Prince

 

[Comrade Ceasefire]

Jimmy not Johnny.

You're probably thinking of the guy from Deep Purple Haze.

 

[Birdjaguar]

Yes Jimi Hendricks virtuoso guitarist.

 

[Bonyduck Campersang]

Spoiler the Iron Tower looming over a desolate landscape, digital fantasy art in the style of Keith Haring :