HOME | DD

Description

A more in-depth experiment with a Stable Diffusion image generator, this time using a fan-favorite Digimon...

Yes. I spent several hours asking an AI to draw Renamon.

BUT also taking time to document each step as I worked out a prompt, then deconstructing it detail by detail, to produce this sequence of results.

Notably, Stable Diffusion produced pretty good results just specifying "Renamon" by name with no further explanation or details from me. After picking a specific RNG seed to chase, among the things that happened during these iterations:

* It started to prioritize an upper-body shot, until I specified that I wanted this to be a full depiction from head to toe. At about the same time, it suddenly decided to draw Renamon naked (...well, moreso than usual), so I had to inform it that Renamon's sole article of wear is the purple sleeves.

* It also forgot about drawing Renamon with arms crossed, but a tweak to the phrasing resolved this.

* I experimented with including a gender (both directly, e.g. "male/female", and indirectly, e.g. "his/her"), which yielded subtle results in body proportions that wouldn't be noticeable outside a direct side-by-side comparison from otherwise the same prompt; Renamon is (by human standards) an androgynous species.

Unfortunately I do not recall the exact phrasing used for the final, full-size image on the right; it was one of a few batches that I deemed "final", but their specific prompts varied, which also means I'm unable to reliably re-produce them (even using their RNG seed).

Look closely and you may also notice that the AI typically struggled with the specific markings on Renamon's upper legs, as well as the metal fence in the background. Stable Diffusion as an algorithm is not well-suited to fine, "fingerprint"-level details like this ... (and for the love of Lorem Ipsum do NOT attempt to make it generate text!)

Ironically, this experiment actually ran longer than it would have taken for me to draw Renamon myself and get everything right on the first try. AI can produce results "faster" than a human but at what cost?

Description

A more in-depth experiment with a Stable Diffusion image generator, this time using a fan-favorite Digimon...

Yes. I spent several hours asking an AI to draw Renamon.

BUT also taking time to document each step as I worked out a prompt, then deconstructing it detail by detail, to produce this sequence of results.

Notably, Stable Diffusion produced pretty good results just specifying "Renamon" by name with no further explanation or details from me. After picking a specific RNG seed to chase, among the things that happened during these iterations:

* It started to prioritize an upper-body shot, until I specified that I wanted this to be a full depiction from head to toe. At about the same time, it suddenly decided to draw Renamon naked (...well, moreso than usual), so I had to inform it that Renamon's sole article of wear is the purple sleeves.

* It also forgot about drawing Renamon with arms crossed, but a tweak to the phrasing resolved this.

* I experimented with including a gender (both directly, e.g. "male/female", and indirectly, e.g. "his/her"), which yielded subtle results in body proportions that wouldn't be noticeable outside a direct side-by-side comparison from otherwise the same prompt; Renamon is (by human standards) an androgynous species.

Unfortunately I do not recall the exact phrasing used for the final, full-size image on the right; it was one of a few batches that I deemed "final", but their specific prompts varied, which also means I'm unable to reliably re-produce them (even using their RNG seed).

Look closely and you may also notice that the AI typically struggled with the specific markings on Renamon's upper legs, as well as the metal fence in the background. Stable Diffusion as an algorithm is not well-suited to fine, "fingerprint"-level details like this ... (and for the love of Lorem Ipsum do NOT attempt to make it generate text!)

Ironically, this experiment actually ran longer than it would have taken for me to draw Renamon myself and get everything right on the first try. AI can produce results "faster" than a human but at what cost?