Star Wars: A New Hope as Post-Apocalyptic 80's Anime Is Just as Iconic

Certainly, the vast majority of their output remains technically flawed, deep in the uncanny valley, or plain creepy, so what you see posted on the net is the best 5% (at best) of images.

There also are severe restrictions on composition of images – neural networks can draw faces well, bodies not so much, and struggle even more with dynamic poses and multiple characters. To get decent results you often need a neural network trained for a specific style.

But the fact remains, that now almost anybody can create images of their favorite characters and things without learning how to draw (some learning on how to coax a network for best results is still required). And a lot of people started to experiment with this newfound capability.

Probably the most interesting results so far involve redrawing existing images and concepts in a completely different style.

The most noticeable trend of the last months is re-imagining various properties in the style of old-school dark fantasy films (think Milius' Conan the Barbarian and its rip-offs). But there are many more. For example, look at this experiment with reimagining Star Wars as an anime from 1980s, with some post-apocalyptic vibes:

All images were created with the Midjourney neural network (probably the best general-purpose drawing network so far) and put to music by the clip's author. As you can see, the network is unable to keeping its output consistent, be it starship designs or characters' eye and hair color, and still has big trouble drawing hands.

(Check the number of Obi-Wan's fingers on the fourth image for the most obvious example, but really, just look at characters' hands whenever those appear in the image).

It is also very likely that the clip's author used actual images of the characters as prototypes, at least in some cases. But the results are still quite impressive. Starships came out especially well – there are plenty of actual sci-fi book covers from 80s and 90s with worse designs.

With further improvements, like the ability to memorize a design and reproduce it with different poses/angles (so far neural networks are only good at creating variations of a design with incremental changes), this technology can become very useful.