Hacker News Reader

Comment by IAmGraydon | Hacker News Reader

I'm wondering a good way to create 2D sprite sheets with transparency via AI. That would be a game changer, but my research has led me to believe that there isn't a good tool for this yet. One sprite is kind of doable, but a sprite animation with continuity between frames seems like it would be very difficult. Have you figured out a way to do this?

View on HN

I think an important way to approach AI use is not to seek the end product directly. Don’t use it to do things that are procedurally trivial like cropping and colour palette changes, transparency, etc.

For transparency I just ask for a bright green or blue background then use GIMP.

For animations I get one frame I like and then ask for it to generate a walking cycle or whatnot. But usually I go for like… 3 frame cycles or 2 frame attacks and such. Because I’m not over reaching, hoping to make some salable end product. Just prototypes and toys, really.

I was literally experimenting with this today.

Use Google Nano Banana to generate your sprite with a magenta background, then ask it to generate the final frame of the animation you want to create.

Then use Google Flow to create an animation between the two frames with Veo3

Its astoundingly effective, but still rather laborious and lacking in ergonomics. For example the video aspect ratio has to be fixed, and you need to manually fill the correct shade of magenta for transparency keying since the imagen model does not do this perfectly.

IMO Veo3 is good enough to make sprites and animations for an 2000s 2D RTS game in seconds from a basic image sketch and description. It just needs a purpose built UI for gamedev workflows.

If I was not super busy with family and work, I'd build a wrapper around these tools

That’s great info. Thank you. I’ll give it a shot tomorrow.

I’ve been building up animations for a main character sprite. I’m hoping one day AI can help me make small changes quickly (apply different hairstyles mainly). So far I haven’t seen anything promising either.

Otherwise I have to touch up a hundred or so images manually for each different character style… probably not worth it

I dont use AI for image generation so I dont know how possible this is, but why not generate a 3D model for blender to ingest, then grab 2D frames from the model for the animation?

Because, uh, literally everything. But the main reason is that modeling is actually the easy (easiest) part of the workflow. Rigging/animating/rendering in the 2D style you want are bigger hurdles. And SOTA AIs don't even do modeling that well.

Reply on news.ycombinator.com