The simulation method used is Material Point Method, a technical that is used in engineering simulation but hasn't really been used much in Visual Effects. This will likely change now that there are these impressive results.
The one issue I've seen people in the industry mention about this paper is that the the simulation times are quite long.
Agreed, very nice.
I've heard the speed issue mentioned as well, although slow compared to what, I wonder? I don't see offhand any fundamental reason why it would be dramatically slower than, say, a standard FLIP-based fluid sim. It amounts to advecting Lagrangian particles combined with solving the elastoplasticity equations on a Cartesian grid, with a snow-specific constitutive model.
There min/frame in the included paper chart range from 3.8m/frame to 25.8m/frame, which is pretty slow for 2013 at the simulation sizes they give.
It is too bad they didn't break out the individual steps to show where the time was spent. Could simply be unoptimized code. I notice there is a restriction on the time step to be less than 0.5 x 10^-3 s, not sure if that is the main difference.
0.5 * 10^-3s is 2000hz, so 20 iterations per frame, assuming they render at 100hz, that seems a bit on the high side.. does that mean they don't have a good way of interpolating the physics effects, does that happen a lot in these kinds of simulations?
I work on a real-time multi-player game, and our physics step is actually larger than our framerate, we interpolate the rest, I guess that's quite the opposite of this :)
In the case of the simulation, the question is whether running the simulation at 24Hz or 48Hz would result in inaccuracies that require them to run it at 2000Hz.
I dont know anything about rendering or simulations, but would this method be possible for real time rendering? or is it only applicable for offline rendered stuff (like movie sfx)?
I know very little about it too, but I think advances in video game visuals have shown that it's really just a matter of time. Check out Uncharted 3's real-time dynamic water effects: http://www.youtube.com/watch?v=BhzMR-vxYIk (Skip to about 2:50)
I noticed some moments in that video where a flashing controller button appeared in the corner. Does that mean the player has to mash the button to continue? I hate games that do that. Why do game makers do that?
I suppose it makes you feel more involved, rather than just watching a bunch of cool cutscenes that you have nothing to do with (common in stylish Japanese RPGs like Devil May Cry, etc)
I don't know how complex the computations are, but if it's close to the computing required by standard loose-accuracy water fluid simulation, we could certainly see it in-game, definitely possible even for this gen of consoles. I speculate as water simulations is quite doable even on 5 year old hardware, though tends to be inaccurate. I remember seeing some Nvidia demos showcases things like softbody destruction and collisions a few years back too, so that also could translate.
There are also some really rough tessellation techniques in use to simulate ground snow walking in games out now, like that Assassins Creed with the native americans.
If we assume 5 minutes per frame then even without optimizations we should be able to render 25 frames per second in around 15 years (assuming Moore's law continues to hold)
Frequency increases haven't held. Strictly, Moore's Law is about transistor density - and it's holding so far.
Relevant here is the corollary, that transistors per unit cost doubles every 1.5 years (also held), which in 5 years is approx x10 (10.0793684), so x1,000 in 15 years.
Assuming this simulation is largely parallelizable (likely, because matter acts locally), then for whatever the hardware it ran on cost, in 15 years it should take 5 * 60/1000= .3 seconds per frame (not 0.04s, for 25 fps). In 20 years, it would be .03s (33 fps). But consider that fakey shortcuts that look OK will be assiduously sought.
Moore's Law (and especially this corollary) have a few years left, but maybe not 15 years... However, I'm optimistic about another technology taking over (as has happened previously). There's plenty of scope, considering what brains and cellular machinery can do. (I think the quantum stuff is BS. Probably.)
If you know what is Cauchy's stress tensor, know about finite element analysis, and grad classical mechanics, its understandable. If you don't know about that, you will not learn it in a demo...
I don't know anything about graphics and simulation, but I'm always impressed with these SIGGRAPH demos. Here's the demo trailer for the papers from this year: http://www.youtube.com/watch?v=JAFhkdGtHck (3 min)
It looks pretty good, but still, I think some of these examples look really foamy, and not like real snow. Not all physical behaviour is yet replicated I think.
Of course not, its only a simulation of a model. The data necessary to store all physical attributes including every atom and sub-atomar particles would not fit on any hard disk in the world. The result is nonetheless pretty amazing.
Yeah, the paper talks about how to vary the parameters to simulate powder, wet snow, slush, etc. These techniques were developed for the movie Frozen, and to my eye (I grew up with snow too) they did a great job of simulating all manner of snow and ice.
Yea, powdery snow isn't all that hard to simulate (relatively) I believe. The snowflakes are just particles that are acted upon by an outside force( wind usually). I live and grew up in New England, and the snow in Frozen looks completely realistic.
It's amazing the amount of work put into these movies. Most moviegoers won't realize what it took to simulate snow realistically, but the would have noticed if it was wrong.
I'd like to see some simulations of slab avalanches...The stuff with wet sticky snow looked okay but I think they missed the ability of even low density snow to act cohesively as slabs [1] (which leads to some of the most deadly avalanches). Need to review their math but maybe they only account for particle adhesion and not more crystalline structure (which would obviously make the simulation more expensive).
The woman walking (post holing is the technical term) looked off for the same reason...the snow didn't look slaby/crusty enough to act like it was.
I was talking to my father in law (head of particle research for P&G) about particle simulation for research purposes and he was explaining the complex challenges with behaviour of interactions. It amazes me that even at the mathematical level, a post-doctorate doctor of chemical engineering with a team of physicists/computer scientists/mathematicians/chemists are still facing challenges which Disney seem to have solved - glad to see they are properly releasing this research!
This process yields fantastic, realistic visuals, but I bet you'd find it does not yield the scientific information they're chasing. It's sort of like how the 'standard' linear static finite element method assumes insignificant displacements and no plastic yielding of materials - as soon as you need to deal with any sort of significant level of displacement or distorsion of the material in question, a linear static solver isn't capable of giving you accurate results.
With displacement, it tries to fake it by extrapolating the linear displacement and assumes that applies over the entirety of it, when in actual fact as soon as it displaces, the strain field in the material will change and it needs re-solving.
Except that, if it was what they wanted, then it's hard to explain the walking character at 3:30 that gets his legs almost trapped in the soft heavy melting snow.
There's just no way that you'd get that kind of effect with that kind of snow. It looks more like the kind of thing that you'd get from non-uniform snow that has a crust on top (and that kind of snow is not sticky and not melting).
What i really like about this is that instead of keeping it under wraps and calling it a "trade secret", they openly discussing their technique and releasing it for the advancement of the field. Not only did they write a paper about it, but made it open access! (as opposed to traditional journal publication)[1] So kudos to their effort!
[1] Though this IS disney we're talking about and they certainly don't need journals for getting an audience for their papers
Yep, but I'd qualify that: The demo seems to show wet snow (snowball snow) only. There are a few other types that happen around here, like dry snow, and crystallized snow. Dry snow is light fluffy stuff that doesn't pack. And crystallized snow is made of what feels like little daggers, behaving almost like ball bearings. But still, their result is brilliant!
Perhaps the link should be http://disneyanimation.com/technology/publications/55 (note the added /55) so it goes directly to the post in question, even if the page is updated (permalink was obtained by clicking the Twitter icon).
As a CS major this fascinates me to no end, but reading the paper the physics are a bit much for my brain to handle. Can any physics-inclined people here recommend me some good books/papers so I can learn more and hopefully make more sense of this?
60 comments
[ 3.4 ms ] story [ 172 ms ] threadThe simulation method used is Material Point Method, a technical that is used in engineering simulation but hasn't really been used much in Visual Effects. This will likely change now that there are these impressive results.
The one issue I've seen people in the industry mention about this paper is that the the simulation times are quite long.
https://en.wikipedia.org/wiki/Material_Point_Method
It is too bad they didn't break out the individual steps to show where the time was spent. Could simply be unoptimized code. I notice there is a restriction on the time step to be less than 0.5 x 10^-3 s, not sure if that is the main difference.
I work on a real-time multi-player game, and our physics step is actually larger than our framerate, we interpolate the rest, I guess that's quite the opposite of this :)
I think you meant to say 5:20 ^^
There are also some really rough tessellation techniques in use to simulate ground snow walking in games out now, like that Assassins Creed with the native americans.
http://lesterbanks.com/2012/03/10-great-examples-of-naiad-in...
It depends on the accuracy of them and the resolution / speed they're going to get rendered at.
Assuming this simulation is largely parallelizable (likely, because matter acts locally), then for whatever the hardware it ran on cost, in 15 years it should take 5 * 60/1000= .3 seconds per frame (not 0.04s, for 25 fps). In 20 years, it would be .03s (33 fps). But consider that fakey shortcuts that look OK will be assiduously sought.
Moore's Law (and especially this corollary) have a few years left, but maybe not 15 years... However, I'm optimistic about another technology taking over (as has happened previously). There's plenty of scope, considering what brains and cellular machinery can do. (I think the quantum stuff is BS. Probably.)
The snowballs colliding is extremely impressive.
That being said, the visuals do indeed have a "wow" factor to them.
[1] https://disney-animation.s3.amazonaws.com/uploads/production...
Great work for where they got to, but it's not quite right yet.
I'm sure Disney et al will find ways to distract you enough not to notice.
I'm far more interested in seeing snow falling, especially with (for example) a street-lamp contrasting it. That would be something to marvel at.
It's amazing the amount of work put into these movies. Most moviegoers won't realize what it took to simulate snow realistically, but the would have noticed if it was wrong.
The woman walking (post holing is the technical term) looked off for the same reason...the snow didn't look slaby/crusty enough to act like it was.
[1]http://2.bp.blogspot.com/-4pGis30qlj0/TwcAGuSz8fI/AAAAAAAAAW...
With displacement, it tries to fake it by extrapolating the linear displacement and assumes that applies over the entirety of it, when in actual fact as soon as it displaces, the strain field in the material will change and it needs re-solving.
It's a good tool, but it has its limitations.
There's just no way that you'd get that kind of effect with that kind of snow. It looks more like the kind of thing that you'd get from non-uniform snow that has a crust on top (and that kind of snow is not sticky and not melting).
[1] Though this IS disney we're talking about and they certainly don't need journals for getting an audience for their papers
There's a lot of competition, but still there's a lot of open stuff like OSL, OIIO, OCIO, OpenSubDiv that multiple facilities contribute to.