AI Sprite generation

[Michy]

Hello community, I want to show you a tool that I added to my Aseprite recently that serves to generate sprites with artificial intelligence using Stable Diffusion as a base.

Recently searching between forums related to Stable Diffusion I found a model called Retro-Diffusion and the truth is that it is helping me a lot to create new sprites and ideas, it is true that it does not serve much to generate characters but to generate (items, armor, runes, tiles, etc) is pretty good.

I comment on this model because I know that many people would be useful, the model is paid and is used along with Aseprite.

I show you examples:

- Prompt window:


Result:

Spoiler: Result

You can also use "image to image"

Spoiler: Image to image

This is a "sword" 🤣 supposedly


Prompt:


Result:

Spoiler: 50 Steps

What do you think about it?

 

[samco]

It sucks

 

[Rexxar]

I don't think this model would have enough training data to produce world objects that fit the Tibia perspective (which is not Isometric btw).
So you would be limited to making just varying item sprites.

Also it looks like the model relies on you procuring or producing a base image for it to use as basis, so just typing something into the text box will most likely just produce garbage.

Plus I don't know if I'd want to support a model author whose entire "customer showcase" is a bunch of NFT scams.

 

[TibiCAM]

Name reminds me of @Asprate haha

 

[lexus21]

@Michy what .ckpt u use?

 

[Michy]

@Michy what .ckpt u use?

I buyed from the Retro-Diffusion guy

 

[Onimusha]

I buyed from the Retro-Diffusion guy

could you share with us?

 

[Night Wolf]

whenever you think of text -> transformers
whenever you think of images or generation -> GANs
DALL-E is a transformer because it needs to interpret language and translate into something, so the focus is on understanding correctly rather than actually providing good quality results.
AFAIK VQGAN should be a best bet, but you'll need to study and train the model.