Interesting premise, but unfortunately the results are not. There's nothing in the designs of the Pokemon that ties them to their connected country, so they all just end up looking generic.
It's a been couple decades since I first saw pokemon, but is this what they look like now? Most of them kind of look like the same 2-legged, mammal thing in different primary colors.
Some are two legged, others are 4, still others have 6 limbs like centaurs, and even so others have a 4 legged creature and a two legged TV stand thing following them around.
But I suspect their similar texture is because stable diffusion was trained on the official artwork for Pokemon rather than their in-game sprites/models, though that is just a guess.
Honestly, these monsters don't look much like Pokemon to me. But maybe cause I've played a good chunk of Pokemon.
I think its impressive that Stable Diffusion can make such "artwork", but its not quite "Pokemon". There's something terrible going on with Stable Diffusion's generation of eyes here, I don't know quite what, but its like Stable Diffusion doesn't understand "anime perspective" at all.
The eyes are often "bouncing" between 3/4ths perspective and straight-at-the-camera perspective, sometimes within the same generated monster. The algorithm is also trying to generate a 2nd face in the bottom-right of a lot of these pictures, for some reason.
I am incredibly impressed with this post and with the novel results of Stable Diffusion already.
The result for Ivory Coast
looks like the latest+ dark n edgy brooding Pokemon for ash and company to love on and teach that life is not all grimdark after all, some 5 or so episodes after Ash gets to a new region and re learns type matchups for the Nth time.
A bunch of these (eg, Cameroon) have a weird extra limb, which is basically consistent with my experience of Stable Diffusion: it frequently produces images that are physically impossible, architecturally nonsensical, etc.
I'm not sure other models are better. I don't think they are. Seems like a big problem that won't be easy to solve.
Ever since I saw your pokemon fine tuning, I've been looking forward to trying it myself ! I find your results to be convincing, as a long time pokemon fan !
I made an account to say thank you for sharing your fine tuning code for others to use !
22 comments
[ 6.2 ms ] story [ 73.4 ms ] threadBut yep, I do hope to improve this. (And if I don't, I'm sure someone else will.)
[1] https://github.com/sradc/each-country-as-a-pokemon-stable-di...
Vatican City is pretty good.
It is sort of interesting that United States and US Virgin Islands both have blue bodies and red paws.
But I suspect their similar texture is because stable diffusion was trained on the official artwork for Pokemon rather than their in-game sprites/models, though that is just a guess.
https://bulbapedia.bulbagarden.net/wiki/Mimikyu_(Pok%C3%A9mo...
https://bulbapedia.bulbagarden.net/wiki/Greninja_(Pok%C3%A9m...
https://bulbapedia.bulbagarden.net/wiki/Buzzwole_(Pok%C3%A9m...
https://bulbapedia.bulbagarden.net/wiki/Corviknight_(Pok%C3%...
------
Honestly, these monsters don't look much like Pokemon to me. But maybe cause I've played a good chunk of Pokemon.
I think its impressive that Stable Diffusion can make such "artwork", but its not quite "Pokemon". There's something terrible going on with Stable Diffusion's generation of eyes here, I don't know quite what, but its like Stable Diffusion doesn't understand "anime perspective" at all.
The eyes are often "bouncing" between 3/4ths perspective and straight-at-the-camera perspective, sometimes within the same generated monster. The algorithm is also trying to generate a 2nd face in the bottom-right of a lot of these pictures, for some reason.
The result for Ivory Coast looks like the latest+ dark n edgy brooding Pokemon for ash and company to love on and teach that life is not all grimdark after all, some 5 or so episodes after Ash gets to a new region and re learns type matchups for the Nth time.
+next in line after Mewtwo and Lucario
Ukraine is kind of cool with a custom articuno vibe
I can't see any logic in Venezuela.
I'm not sure other models are better. I don't think they are. Seems like a big problem that won't be easy to solve.
I made an account to say thank you for sharing your fine tuning code for others to use !
I agree that it produces some incredible results, and it's awesome that justinpinkney open sourced it!
[1] https://github.com/LambdaLabsML/examples/tree/main/stable-di...