4 ms·
>And why is it so damn hard to get clothing to actually fall on the body right, instead of always looking like a stretchy foam skinsuit? Mostly because it is a
by jayd16 7y ago
>And why is it so damn hard to get clothing to actually fall on the body right, instead of always looking like a stretchy foam skinsuit?
Mostly because it is a stretchy foam suit. A model is deformed and translated based on animated bones in the model. Essentially each vertex is tied to a small number of transforms matrices(bones) with multiplier/weight. Animations are updating the transform matrices over time. You only get so many bones per model and per vertex and because of this clothes moved inline with other parts of a model like skin because its using the exact same bone.
I'm not sure where we're at in terms of increasing bone counts or independently animating clothes on a character. As bone counts increase, the amount of matrix data that must be sent to the GPU every frame increases.
- random32840 7y agoI can see deep learning shaking up dynamic cloth animation dramatically in the not too distant future. It's already happened for fluids.
- archi42 7y agoFunnily, I wasn't amazed by the fluid in the demo. It seemed much too thick (more like oil than like water).
- faitswulff 7y agoThe fluid was really unrealistic and they panned away from it quickly.
- random32840 7y agoThat's not a great example of what's possible. Research-land has some wild demos: https://www.youtube.com/watch?v=i4KWiq3guRU https://www.youtube.com/watch?v=i4KWiq3guRU
- jayd16 7y agoHow so? We can already render really nice cloth offline. The issue is the runtime cost of sending that animation data to the GPU. Are you suggesting building some kind of GPU side physics protocol that takes material properties, and existing bone data and fills in some of the blanks? I suppose you could start to push bone motion vectors and interpolate some things.
- random32840 7y agoComputing & rendering all on the GPU without having to query back to the CPU at all seems a natural fit to me, but I don't have a deep technical understanding of it. My main thought is just that DL reduces the amount of computate you actually need by allowing you to approximate rather than derive things like cloth physics, how that's eventually integrated seems an implementation detail.