I asked Stable Diffusion 2.1, trained on two different datasets, to draw some glowing pokémon. I specifically picked pokémon that non-players are most likely to recognize. As it turns out, this was a good choice because Stable Diffusion struggled to accurately depict even some of these more familiar pokémon. It did pretty well with Squirtle and Pikachu. But it had a hard time with Bulbasaur, Charmander, Mewtwo, and to some degree Eevee. For each one, the result from the older training model is shown on the left and the result from the newer training model is shown on the right.
1. Bulbasaur
4. Charmander
7. Squirtle
25. Pikachu
133. Eevee
150. Mewtwo
And, just for fun, here are some of the results that weren't sufficiently recognizable as the target pokémon:
Bulbasaur Fails
Charmander Fails
Eevee Fails
Mewtwo Fails
These illustrations were drawn using Stable Diffusion 2.1.
