Hacker News Reader

Comment by tshadley | Hacker News Reader

Dreamfusion -- text-to-3d seems like it might be useful here eventually: https://dreamfusion3d.github.io/ (once successfully opensourced, see https://github.com/ashawkey/stable-dreamfusion)

Rigging also looks like it could have a decent AI/DNN solution: https://arxiv.org/pdf/2005.00559.pdf

View on HN

Once there is a good 3D model generator I’ll probably have to quit my job and start making games again. That’s incredibly time consuming to create even low poly models for someone who is not very good at it. I remember trying to make a moose and after a few weeks of time it was somewhat passable but with no animation or rigging

You can just pay an artist. Afaik Jon Blow did not make any of the art for Braid.

Finding an artist who is (1) capable, (2) reliable and (3) available is no mean feat. I can totally understand why someone would prefer to spend hours (if not days) using an AI tool.

That’s the problem - I don’t have the money to pay someone to work full time for a year or two

Not everyone can or wants to- maybe they want to be the artist through AI

You gotta use sculpting tools like ZBrush / Substance Painter or 3d Coat to do make models like a low poly mouse.

I would NOT recommend sculpting tools for making low poly. That's nest done drawing a basic reference image and then manually drawing/shaping polygons to get the desired look. Low poly is a lot like pixel art.

I have used sculpting (in blender) however, for making more organic 3d printable stuff, since fusion360 absolutely sucks for organic shapes. With a proper drawing tablet sculpting in blender is so much fun tbh.

Rigging has come a long way with AccuRig and Cascadeur in 2022, approaching diminishing returns.

When I asked for a squirrel with a hoodie, it gave me one with two heads:

https://i.imgur.com/UIBXXvj.png

We're definitely at the point where these tools can do amazing things, and yet they still manage to make bizarre defects in the output that no human would ever imagine.

It’s almost like AI takes malicious compliance to a whole new level.

I've played around with these mesh inference methods, and while I certainly value and respect the progress that has been made by people working in this field - Something feels fundamentally off about it - as if they aren't being quite ambitious enough.

Consider shape-net - while being able to generate samples from the shape-net distribution would be a very impressive achievement, and these would certainly be useful as "assets" (decorations in a game scene), they don't have any of the real structure of these objects that would allow real interaction. They are effectively just inflatable decoys compared to real objects.

In short, "AI" isn't going to replace the artists, it's going to massively amplify their productivity and we don't know of any limit to consumer demand for realism and rich interactivity in games. If anything the number of game artists is going to increase, only the bar will be raised on the quality of their product.

I'll be curious to see if text-to-3d is better than text-to-2d with some other 2d-to-3d process.

I guess text-to-3d offers the potential of automatically creating entire worlds with very little human/artist intervention. But text-to-2d matches current processes with more chances for an artist to intervene, and also has more training material for art styles.

It'll probably start with 2d -> 3d because a lot of work has already been put into tools to automatically take images and convert them to 3d. Ultimately direct to 3d will work better, but we will probably need more 3d data for it to reach it's potential, so 2d -> 3d might last for a while. In the interim, people will do all sorts of hacks on top of 2d -> 3d to reduce the amount of required human intervention.

Reply on news.ycombinator.com