Genie: Text to 3D in Seconds
lumalabs.ai
lumalabs.ai
Why do companies do this? Is it a development shortcut or something?
Yeah for sure. I guess one advantage is that since it's already in a chat, people can react and talk about it as it's happening.
One time I participated in a small local event that required submitting photos via tagged Instagram posts. I was annoyed at that, until I realized it completely eliminated the need for any technical expertise for the organizers.
- single sign-on
- abuse protection
- has sign-on flow for later if you want to migrate
- someone else hosts assets
- existing account flow means low friction to join for existing account holders (common in the space)
- allows users to see other users at same time as use product
- bot easy to spin up and async functionality is inherent
It's annoying but I can see why its happening
There would still have to be hosting, moderation controls, high availability, low latency, etc. etc. engineering that "someone" would have to handle and own (ideally decentralized / p2p), but I believe it could be solved if ever incentives could align and a will could arise.
It's annoying to have to do this for every server though, I agree.
edit: worked on my phone
and not so good examples:
https://lumalabs.ai/genie?one=15889f37-f001-43ca-bfbf-cd75d2... https://lumalabs.ai/genie?one=b7352636-2203-4ec2-9b5b-d1dd80...
In the initial preview models look less lumpy and overall much better. So I guess the transformation from 3d pointcloud/gaussian splatter or whatever it uses to 3d meshes isn't great.
This is going to be a longer slog to solve than image generation (though I'm sure we'll get there in a few years).
If you view this link https://lumalabs.ai/genie?s=1&d=-4, do you see the models more clearly?
If you _don't_ I'd be curious if any errors are reported in console
How long would you expect a single graphic to hold your attention before you can deem it impressive? Five minutes?
In good hackernews style, I got downvoted into oblivion for just a simple question, but I guess that's to be expected these days.
Do you get any errors in the js console? Are you on Destop or phone?
We start with a point cloud generated from the model which has color and normal and render each particle as a little hexagon with metallic and iridescent shading. Each particle is given a position and velocity in a texture. Each frame we do a physics step that writes to these textures with fragment shaders. We sum the forces into the velocity texture:
- mouse interaction force
- flow force from time varying curl noise
- spring force to return to original position
- air resistance
last we do a euler step to update position texture. There’s masses of free parameters to control forces and materials, the animation is created by changing parameters per-particle over time
Tech: three.js, GLSL & TypeScript
Tip: Left click to push apart, right click to pull together!