I tried to see how far I could go with AI for mecha design. It was pretty insane.

An experiment in endlessly tuning prompts to create cool robot concept art.

Below are robot studies created by giving the AI plain text prompts. (The lighting and panel lines are purely my own preference, so it's not like it can only make robots of this style.)

From OpenAI's DALL-E 2, Disco Diffusion, and MidJourney, to trying out my own StyleGAN and VQGAN systems, as well as CLIP-based ones...

After trying various things, I've figured out good parameters and ways to give commands. It really is easier to get high accuracy when you study by looking at the implementation and code yourself to understand the internal behavior of the algorithms.

Even with the same command, different results come out depending on the internal implementation, the libraries used (or assumed to be used), and the evaluation functions. This kind of know-how seems likely to be hidden away as "spells" among top professionals.

From here on, it seems possible to infinitely expand quality and range by using multi-connections, such as "post-processing things made by AI with another AI."

Combining words unrelated to architecture or robots seems to make it easier to generate unseen shapes that are less likely to look like copies.

Study for cover image

Generating a landscape format for a note cover

Notes on AI weaknesses and future challenges

Since each one is generated independently, it seems to struggle with things like generating "different angles of the same character." (It's not impossible if you get creative with the command prompts, but it requires a lot of know-how).

In the near future, it's possible that "rough concept art for capturing a mood" will move out of human hands. On the other hand, fully controlled works that focus on specific issues or subjects can still only be created by humans.

Discerning which skills will be replaced by AI and which skills will remain valuable even with AI is going to be a challenge for every creative profession over the next decade.