Midjourney - AI art

[Rroff]

they almost look like real people!

Looks like a fair few likenesses of famous people composited into the faces.

 

[Derek W]

from this

to this

 

[tang0]

4o image by far the best model yet

 

[Rroff]

from this

What is interesting even the more basic models can get pretty close to recreating that scene just using a basic description as input. Though for some reason they persistently want to face the motorbike the other way even if you put in specific prompts.

While a bit further from likeness I thought this one was kind of cool - like the cover art from some 90s video game:

Spoiler

 

[nine_tails]

from this

to this

Hot damn, thats one epic picture.

 

[Efour]

Archer got really old lol

This is on my list of things to do/dabble with.

 

[Mr Jack]

Though for some reason they persistently want to face the motorbike the other way even if you put in specific prompts.

Try getting it to draw a full glass of wine.

 

[Rroff]

Try getting it to draw a full glass of wine.

How full? most of the models seem to want to do something between empty and half full but some variations are nearly to the brim like this weird and wonderful creation :s

 

[Mr Jack]

How full? most of the models seem to want to do something between empty and half full but some variations are nearly to the brim like this weird and wonderful creation :s

That's amazing.

They can't do full glasses, no matter how much you ask, because of how they regurgitate their training data and the lack of photos of full glasses in that data. I thought it was an interesting insight into how they work and where the limits lie.

 

[Rroff]

That's amazing.

They can't do full glasses, no matter how much you ask, because of how they regurgitate their training data and the lack of photos of full glasses in that data. I thought it was an interesting insight into how they work and where the limits lie.

I tried fudging it with inputs like "overflowing" but the output are still 30-60% full glasses LOL. Does show the limits of these models.

Bing's best understanding of overflowing:

 

[Mobius 1]

correct, I can’t seem to get anything past edit* 65% full.

 

[Rroff]

Going quite abstract I can get some good results actually - you need to start describing wine by other terms.

EDIT: Yeah if you describe what a wine glass is then tell it to fill it with a red bubbly liquid it can manage it - what it doesn't understand is a "full wine glass".

 

[Mr Jack]

Going quite abstract I can get some good results actually - you need to start describing wine by other terms.

Get it to fill that to the brim.

 

[Mobius 1]

Get it to fill that to the brim.

Last attempt.

 

[McGray]

Last attempt.

We have a winner.

 

[Mobius 1]

Using ChatGPT 4o model. Got another one ever so slightly fuller too.

Edit

 

[Rroff]

Last attempt.

About as good as I can get it with a lot of fudging, don't think I want to spend all day trying to get that last 0.5%.

 

[Mobius 1]

About as good as I can get it with a lot of fudging, don't think I want to spend all day trying to get that last 0.5%.

That was my 4th attempt. Seems like the new 4o model is better at handling specifics.

 

[Ayahuasca]

Close but no cigar:

 

[dowie]

Try getting it to draw a full glass of wine.

EDIT: Yeah if you describe what a wine glass is then tell it to fill it with a red bubbly liquid it can manage it - what it doesn't understand is a "full wine glass".

Actually, it kinda does understand - it knows that a wine glass shouldn't be filled all the way up, 1/3 to 2/3 is a full pour.

Obvs in the second pic it isn't quite to the brim but it's basically filled the glass, in the first pic it's drawn a wine glass filled as it should be albeit a bit more than the 1/3 it says is correct.

Edit a simply prompt telling it that the wine isn't quite full to the brim corrects that second pick: