It's not meant to look like the material snow. What's shown here is a physical simulation of snow, not a visual one. We can already make a surface look realistically snowy with the right properties, mainly subsurface scattering; the tricky part is the physics of it all. This simulation looks like shaving cream because it's a like a technical paper, it's meant to make the simulation's details clearer rather than make the snow look realistic.
@@SgerbwdGwyn That's where I'd disagree. Having dealt with physical snow in every variety from slushy to fluffy to grainy...too much of it in fact (yay Canada)...this does not behave like any of them. It behaves like foam. Yes, the renderer could be stronger in the scattering/refracting department but it still stands that this looks more like a foam slip and slide than snow being pushed around. Perhaps it needs more 'fluid object interaction' to push back harder on the scene objects... The glaring part is where it drips off the rooftops.
@@frollard I agree. Some of these were close to convincing when it comes to very dry crystalline snow. The single tire rolling for example. But others, while impressive no doubt, had some very uncanny valley behavior going on. The snow plowing was way off. The particles needed more stickiness I think? The guy with the snow angel also threw me off - I think it was the movement of the snow above/behind him. It was moving too....elastically? Never the less, still by far the closest I've ever seen to even trying to replicate proper snow in a simulation. Neat stuff.
@@frollard I agree! Not to mention the simulation lacks the layering that snow takes on, like the top layer that melts and refreezes into a shiny hard crust of ice with it's own wildly different properties or how different moisture content snow will stick to itself and different snows in different ways, IME fresh snow pack is rarely homogeneous as shown here, the very top and bottom layers are quite different, especially where new snow falls over already compressed snow. It's not clumpy enough to be wet snow and not sparkly or powdery enough to be, well, powder. Too much foam, not enough ice
4:34 i mean yeah, mega = million, so 300MB would mean each particle takes 300B, which i think is a little much, but i suppose there’s temperature, location, bonds with other particles, and other “hidden” data that’s required for the simulation
@@donpercent Well, the simulation needs to take more things into account like the varying degrees of ice stiffness/age, since the compactness trough pressure is variable according to snow age and overall temperature
@@kinkias I think the simulation needs to take into account the actual structure of ice crystals in snow. If they represent individual tiny snow flakes as little spheres it can never really look right. The problem is that to simulate reality nearly perfectly takes so much detail and information that computers can't actualy do it. The over night rendering would turn into a month long rendering that ultimately crashes your computer. Lol
@@Redd56 might be able to run the sim but you wont be able to display it in any meaningful detail. which makes the sim quite useless, for realtime applications.
An important part of our original Disney paper is rendering the snow as a volume. It makes a huge difference. In principle you could particle density estimation to get a density field from the sph particles and then do a volume render from their method as well. It would look a lot better. This paper is more about the dynamics than the look.
I got a question: Going from hours to minutes is great, but "is the increase in performance comes from a more efficient algorithm?" or "is the increase in performance comes from using newer processors?" Were the referenced hours and referenced minutes benched on the same hardware? If not, then it would seem dubious to have the information included. Anyhow, the quality has improved no doubt. But, how has the performance really improved? Love the video, thank you for your work.
In this case the answer is obvious because the change in render time is in the order of several magnitudes and hardware obviously hasn't improved even close to that.
You can pretty much guarantee they weren't done on the same hardware, but the hardware used here is specified in the paper: a 16-core 3.1 GHz Intel Xeon E5-2687W workstation, which is about 6- year- old tech but still much faster than any processor you're going to see in a modern PC. In general, the time frames given in articles like this are for a sense of practical time commitment, and not for comparison with previous work. If there was a serious question about the relative efficiencies of the two approaches, then they would compare them side-by-side, but in this case the improvement is extreme and obvious, and therefore unimportant.
@@michaelleue7594 indeed. I remember reading how each frame in a Pixar movie took hours to render and ran across 12-96 cpus/ PCs (server farm). It is good to know if the new software is running on a standard PC as those are the times most close to what you, the viewer, will see. IIUC in each of these papers it is the software that is bringing the vast improvements NOT the hardware. Faster CPUs only show marginal improvements. There is hardware improvements but it comes from having a GPU that supports CUDA and tensor flow (AI related number crunching) that improve the AI algorithms that are mentioned in the papers.
@@FunielAudio Yeah, much faster. It's a workstation chip. The clock speeds are not the main consideration when you're comparing fundamentally different hardware. This thing runs servers.
The way the snow slides/rolls of the roofs also needs some work. i think it should move more like a whole unit instead of rolling down like that. next paper maybe :P
Hopefully a few papers in the future will take care of the various types of snow, and make the snow's viscosity/deformations look less weird/goopy 3:57/look less like grainy like Styrofoam/ look more realistic when attaching to clothes 4:19. Otherwise fantastic work, progress, effort!
They probably used the same friction value for everything in the scene and optimized it for the snowplow. The roofs should have a higher friction value
That is really cool how little memory it uses, but as an FX artist I do snow simulations and I can tell by those renderings that they're using like 10% of the amount of particles I normally use. Im just wondering if that's the extent of detail for that program or if that's just the extent if it being real-time? Either way really cool how efficient it is!
I guess the big majority of the scenes is "static" snow without interaction, so they transformed it into a simple mesh and took care only about the snow close to the interactive part of the scene
Its because its not meant to its just showing the simulation of the particles in actual use for visuals the artist would add shading etc and change variables to make it actually look like real snow.
Very impressive especially for the time it takes to track all the particles, but hopefully they can continue to improve the accuracy and how realistic it is, because it acts kinda funny compared to snow, both soft and dense snow sims act more like sand or foam, and the snow seemed to lack momentum especially with the plow
I feel like this still has a ways to go for realism. The snow looks like it's acting more like a liquid than it should, especially the scene with the vehicle scooping the snow. Just something to look forward to though, can't wait to see the day.
there are many fine achievements with this latest work. i wonder if it would be very difficult to improve the snow plough sequence by leaving a layer of snow still on the road? it is like this in some places in the demo clip, but in others, the road is left clean and pristine, which is not very realistic. it would be great if discolouration could be added to the snow, which soon takes on a grey colour (unfortunately), at least in cities. also, twigs and other debris fall on snow pretty quickly. not strictly speaking snow itself, but still a vital part of making it look realistic.
That might have more to do with the plow and the road both being too pristinely smooth and in perfect contact with each other. when it went over the bump, it did leave some, but i agree it looked a little foamy. Still, very impressive!
Discolouration is hard to accurately simulate, as it involes simulating dust and debris particles in the air interacting with the snow. (Atleast I think so.)
@@meeeshroom no question that the discolouration might be tricky, but for feature film work (which is what i need), things have to look highly believable otherwise the audience spots it and it erodes the dramatic suspension of disbelief. i am sure that with the right algorithms (and training) something highly realistic will be achieved.
@Bronze HD yes, that is a fair point. it would be interesting to see the results if a rough texture were used. if done with 3D geometry that could be expensive on polys, but i have not read the paper - maybe there is a neat and light way of doing it.
The Inuit of the Frozen North which is in Canada, have over 2000 words for snow. Powder snow, wet snow, mushy snow, clumpy snow, frozen snow... you get the picture. Can this snow simulation, simulate wet snow mixed in with lake effect snow? Or should we wait for the next paper?
Thanks for this video but I want to point out that the more complex scenes still run at 8-10min/frame. Which means that a 10sec scene would simulate for 40 hours. (See graph on page 9)
This is going to be difficult. The bunnies were way off, since the snow didn't melt and compress, it crumpled. Also, snow doesn't poor from roofs like that, it sticks to the roof and to the already present layers. The tearing is also weird or largely up scaled. You might as well do thermal capacity simulations because various things will take a long time before a snow layer starts appearing (e.g. on a wet tarmac).
yikes, still a lot of work to get the plow simulation working well - I live in a particularly cold climate, (outside it is currenty sub-zero F) and I can tell you the truck with the plow looked really unrealistic. Hopefully, that was older footage. The simulation essentially looks like viscus water, not snow.
I expect many more papers, because, while being impressive work, this looks and behaves nothing like snow. It looks more like what styrofoam balls would do.
The general shapes of the simmed snow look more like a wet putty or mud. The patterns don’t reflect the grainy, crystalline quality of snow or the way it crunches/breaks apart.
Super video as always, but for someone who live in the middle of a heap of snow at the moment, I have to say it does not look realistic at all. It resembles styroform particles.
How is the "time per frame" calculated? When the results are compared with older papers do they take in consideration the hardware and software upgrades?
Hold up. That blue guy seemed to walk unimpeded by snow. The snow reacted to him, but he did not react to the snow. As I've grown to expect perfection from this channel, this is simply unacceptable 😉
Love these vidoes, even though I am not from this field its fascinating to see the advancements and wonder how much work went into achieving such beautiful results. As an aside: The gamer in me is dying to see how games of the future might apply these in real time!
Good question, Huzzzer! Yes, there are many real applications in the real world. Simulations are already being used to better demonstrate products to customers online. Allowing the customer to virtually hold and get a better look at the product can be improved. Simulating how clothes would look on the customer is making great strides. Also, simulation are increasingly being used in industrial control systems and construction to make realistic views and sounds, and to realistically simulate operation. The industry is evermore embracing the use of actual video game technology to create the simulation because of the improvements in, physics simulation engine components, interface, and costs. These Two Minute Papers videos do an excellent job of showing what is providing some of these advances.
I'm sorry if this is a stupid question: how does the "time per frame" metric account for computational power? I'd imagine that within one paper, all of the measurements were made using the same computer hardware, but how would you compare between different papers from different times?
does anyone else think this looks.. not correct? Examples: At 2:05, a tire won't create styrofoam-like edges, with snow particles drizzling off - it will compact the snow edge, more resembling a knife cut. At 3:54, the ice cream wouldn't split into many pieces with the center staying rigid - it would all collect in a single blob on the ground, with maybe som small splatter.
At 2:41, only very wet snow would come close to retaining its rabbit shape. Very wet snow would not turn into fine powder, as seen when the ears fall off.
Maybe I'm overly critical, but as someone living in the arctic, I see that these simulations clearly deviates from reality to a large extent. So I find it a bit strange when Károly's commentary seem to imply, from the footage alone, that this paper is an amazing achievement. It might still be, for all I know, but I'd prefer if the commentary also acknowledged the weaknesses and limitations.
I remember doing some explosive simulations on my old Intel 2600K back in the days. Simple simulations took over 8 hours, for only 10 seconds. Now we can do this with a CPU that is equal to an Ryzen 3800X: What a time to be alive!
As an incompetent person in simulations I can only tell that it looks like some kind of fluid or foam to me. It might be a huge milestone in simultaions as far as techniques are concerned but for visuals it is just not the best. Regards! Nagy rajongója vagyok a csatornának és a munkásságának, így tovább!
It's not meant to look like the material snow. What's shown here is a physical simulation of snow, not a visual one. We can already make a surface look realistically snowy with the right properties, mainly subsurface scattering; the tricky part is the physics of it all. This simulation looks like shaving cream because it's a like a technical paper, it's meant to make the simulation's details clearer rather than make the snow look realistic.
I don't know how it works exactly, but where do all these simulations end up? I mean, what is stopping to add all these features in unreal engine (for example) right now? Selling that to epic games would create a new generation of videogames. Photorealistic graphic is mostly already achieved for triple A games, but what is lacking is proper physics simulation of nature. Also 3d artists would love this.
I'm halfbrain Artist that can do only simple simulations. I don't understand why it's so good except effeciency. To me it's looks like a slightly frozen foam with slushy cream inside. Does it supposed to have second pass simulation with small particles of snow in cracked areas or something? I understand that it's not production shader for final render. I just complain about lack of my imagination "how it might work in real world VFX".
I was wondering if UA-cam Company could ever develop a client side algorithm, that buffers frames of the video in low resolution, then enhances resolution using ML to result in a genius predicted resolution. This way network will suffer less, the computation is on the client side, get nice no glitching video frames with high resolution. Note: this suggestion if for countries with poor network connection. Any humble comments are accepted!!
Given that half the comments seem to be saying this is the most realistic simulation they've ever seen, and half are saying that no... this doesn't look right, too sticky/foamy/liquid looking.... I'm pretty sure we must have two very different types of snow in the world, and I've apparently only lived in places that get the kind that doesn't look like this. I think this one stills has a ways to go before it looks remotely realistic. Two more papers?
Greetings from Finland. It seems I managed to procrastinate clearing up the snow at the yard till tomorrow, watching Two Minute Papers simulating snow. ;) I have to say these simulations look more like 'wet dry powder' than actual snow in behavior. Material in video is absolutely too fluid and uniform in density, droops right after it's fallen, compacting and solidifying is missing, melting simulation misses the state where snow still holds its shape but wets (I guess it needs more thermal than particle simulation) and so on.
Károly, you need to share your course on light transport simulations with the authors of this paper. Not trying to diminish the authors' achievements in snow simulation but the material choices are just poor. As others mentioned in the comment section, the snow looks more like a kind of shaving foam. On the other hand, Disney's paper uses very believable materials and that makes it harder to compare the papers directly. Can't wait for new video games in snowy environments (few papers down the line, as always).
Looks good, however it seems there's too much focus on the interaction of small individual particles. Snow really clumps when it starts and goes above 0. However this of course depends on the conditions before hand. I'm Canadian heh
It's definitely a leap forward but it's amazing how far it still is from realism. Just shows how much work there is still left for the people joining the field each day. You can see that they don't have the level of dynamic change and variation that snow has. The powder snow was too sticky and the densely compressed snow wasn't sticky enough. They need to add a layer that adjusts the friction and cohesion dependent on compression. Powder snow won't make a good snowball but if you run it over with a car it'll compress very solidly. The snow pushed by the snowplow should turn into hard dense packed chunks as it pushes out. I'm also curious how they deal with partial melt because a big element to how snow acts is partial melts at different stages and interaction with small amounts of melt changing clumping properties.
They react like Styrofoam particles and nearly all are copies in that the single individual object in that all react similarly, they "melt" the same way. You do not take into account the cohesion of the ice crystals and instead treat each particle as it's own entity. Not even powder snow reacts this way. Your presentation looks great btw, sadly, snow does not react like this. At the very least you need to add in some cohesion factor based on temp, to include the ice crystals melting singly or in clumps depending on temp and environmental factors and objects like the car used in this demo. The car was painful to watch as snow does not react that way when driven over. You need to add many new factors and variables. Cohesion due to temp (single entity crystals landing when frozen, melting together slightly, or simply interlocking when cold, their overall weight and reaction when combined etc) and compression ( what happens to that square of snow due to its own weight? Weight with temp gradients, refreezing?) are two that need to be added.
You could have told many thing, but this being snow? No. It strikes me rather like some magic styrofoam that can turn into a sticky gel. it has some strange isotropy to the way it is moving and breaking, the harderning seems like a "oh, stress, lets just randomly combine with the surrounding", and when in contact with other surfaces it turns into a gel. 2:08 shows that really clearly. Just look at the behaviour at 1:30 - yes, it leaves some sort of (strange) tire marks but everything else does not make it seem like snow. By how it looks i kinda also doubt that it can even come near disneys version in terms of behaviour - just snow rolling down a slope and accumulating.
I live in Russia, and from my own experience I can tell that this doesn‘t look like snow at all. More like sour cream or something. Especially on the snow city and car demos. :( Disney’s snow looks a lot better.
20 years later: Dear fellow scholars, today we have great news: The machine learning algorithm can now simulate all possible interactions of quantum particles from scratch in only 4 minutes!
I'm sorry. I live in a country being snowy right now and I can tell you: This doesn't look like snow. More like packaging material or little balls of styropor.
*"7 Years of Progress In Shaving Cream Simulation!"*
Came to comment this. This is stay puft marshmallow man shaving cream sim. It's close but uncanny
It's not meant to look like the material snow. What's shown here is a physical simulation of snow, not a visual one.
We can already make a surface look realistically snowy with the right properties, mainly subsurface scattering; the tricky part is the physics of it all.
This simulation looks like shaving cream because it's a like a technical paper, it's meant to make the simulation's details clearer rather than make the snow look realistic.
@@SgerbwdGwyn That's where I'd disagree. Having dealt with physical snow in every variety from slushy to fluffy to grainy...too much of it in fact (yay Canada)...this does not behave like any of them. It behaves like foam. Yes, the renderer could be stronger in the scattering/refracting department but it still stands that this looks more like a foam slip and slide than snow being pushed around. Perhaps it needs more 'fluid object interaction' to push back harder on the scene objects... The glaring part is where it drips off the rooftops.
@@frollard I agree. Some of these were close to convincing when it comes to very dry crystalline snow. The single tire rolling for example. But others, while impressive no doubt, had some very uncanny valley behavior going on. The snow plowing was way off. The particles needed more stickiness I think? The guy with the snow angel also threw me off - I think it was the movement of the snow above/behind him. It was moving too....elastically?
Never the less, still by far the closest I've ever seen to even trying to replicate proper snow in a simulation. Neat stuff.
@@frollard I agree! Not to mention the simulation lacks the layering that snow takes on, like the top layer that melts and refreezes into a shiny hard crust of ice with it's own wildly different properties or how different moisture content snow will stick to itself and different snows in different ways, IME fresh snow pack is rarely homogeneous as shown here, the very top and bottom layers are quite different, especially where new snow falls over already compressed snow. It's not clumpy enough to be wet snow and not sparkly or powdery enough to be, well, powder. Too much foam, not enough ice
Imagine 2035's new battlefield game including this snow effects.
4:34 i mean yeah, mega = million, so 300MB would mean each particle takes 300B, which i think is a little much, but i suppose there’s temperature, location, bonds with other particles, and other “hidden” data that’s required for the simulation
This is making me laugh at myself because I don't know how snow works since I haven't seen any in real life.
I thought everyone just used smoke sims for avalanches
4:30 помню манку в детском саду так мешал
These scientists should go outside and look at real snow lmao
Still looks like a mix of styrofoam, whipped cream and shaving cream..
Yeah, nothing like a real snow and I dont think this will ever be possible to make it look real.
@@donpercent Well, the simulation needs to take more things into account like the varying degrees of ice stiffness/age, since the compactness trough pressure is variable according to snow age and overall temperature
its fast, but royally bad results.
@@kinkias I think the simulation needs to take into account the actual structure of ice crystals in snow. If they represent individual tiny snow flakes as little spheres it can never really look right. The problem is that to simulate reality nearly perfectly takes so much detail and information that computers can't actualy do it. The over night rendering would turn into a month long rendering that ultimately crashes your computer. Lol
@@dickrichard626 Not yet, but give it a few papers. Better simulations will come.
ah yes frozen liquid
I mean, i guess
@@waldolemmer I know right people just like comments that are verified :D
frozen liquid ice
Technically, almost all solids are frozen liquid.
Ahh
I wish they had some as a direct comparison to the Disney paper
Maybe watch frozen haha
@@beowulf2772 But the speed difference is important too.
I've never been happier for an Icecream to fall.
The physics are amazing, but their shader ruins the final result.. Looks like 1990 CG snow
True but thats probably trivial compared to the sim
I agree that the shader breaks the demo.
@@Redd56
might be able to run the sim but you wont be able to display it in any meaningful detail. which makes the sim quite useless, for realtime applications.
An important part of our original Disney paper is rendering the snow as a volume. It makes a huge difference. In principle you could particle density estimation to get a density field from the sph particles and then do a volume render from their method as well. It would look a lot better. This paper is more about the dynamics than the look.
@@rocksfire4390 a youtube video doesn't need to be real-time and presentation is far more important
No matter how bad things are, whenever Károly ends with “what a time to be alive!” I feel better.
I got a question:
Going from hours to minutes is great, but
"is the increase in performance comes from a more efficient algorithm?" or
"is the increase in performance comes from using newer processors?"
Were the referenced hours and referenced minutes benched on the same hardware?
If not, then it would seem dubious to have the information included.
Anyhow, the quality has improved no doubt.
But, how has the performance really improved?
Love the video, thank you for your work.
In this case the answer is obvious because the change in render time is in the order of several magnitudes and hardware obviously hasn't improved even close to that.
You can pretty much guarantee they weren't done on the same hardware, but the hardware used here is specified in the paper: a 16-core 3.1 GHz Intel Xeon E5-2687W workstation, which is about 6- year- old tech but still much faster than any processor you're going to see in a modern PC.
In general, the time frames given in articles like this are for a sense of practical time commitment, and not for comparison with previous work. If there was a serious question about the relative efficiencies of the two approaches, then they would compare them side-by-side, but in this case the improvement is extreme and obvious, and therefore unimportant.
@@michaelleue7594 indeed. I remember reading how each frame in a Pixar movie took hours to render and ran across 12-96 cpus/ PCs (server farm). It is good to know if the new software is running on a standard PC as those are the times most close to what you, the viewer, will see.
IIUC in each of these papers it is the software that is bringing the vast improvements NOT the hardware. Faster CPUs only show marginal improvements. There is hardware improvements but it comes from having a GPU that supports CUDA and tensor flow (AI related number crunching) that improve the AI algorithms that are mentioned in the papers.
@@michaelleue7594 still much faster? I mean it has a lot of cache but the clocks seem very slow compared to any modern CPU (eg an i7 10700k)
@@FunielAudio Yeah, much faster. It's a workstation chip. The clock speeds are not the main consideration when you're comparing fundamentally different hardware. This thing runs servers.
Very nice, although on some of the examples, it looks more like a foam then snow (like 1:15 or 1:30)
I think that's just rendering not due to the actual simulation
its because the metaball/liquid effect that attaches them
Its just the rendering techniqe,
slap some refraction on that bad boy and you got AAA snow
The way the snow slides/rolls of the roofs also needs some work. i think it should move more like a whole unit instead of rolling down like that. next paper maybe :P
The man laying down in the snow also show very foam like physics.
4:26 when he moves his arms up, way too much snow moves
As I play Death Stranding I always wonder why people aren’t talking about how amazing the snow effects are.
Because everyone talking about: WALKING SIMULATOR! And cant understand, this game is soo good...
@@TeslaElonSpaceXFan,
Amen.
Death Stranding? I checked that out, looks very basic to me. Haven't you seen Red Dead Redemption 2 snow yet? Check that out! That is snow simulation!
There should be awards for these things in games like best snow, best ocean, best woods etc.
@@Kie-7077 agreed
Looking forward to video games being able to do this in the future.
Hopefully a few papers in the future will take care of the various types of snow, and make the snow's viscosity/deformations look less weird/goopy 3:57/look less like grainy like Styrofoam/ look more realistic when attaching to clothes 4:19.
Otherwise fantastic work, progress, effort!
Might be just a question of shaders, these animations are to just show off the physics. But honestly I'm just talking out of my ass here.
To be honest, it looks and feels a lot more like a cottage cheese simulation than a snow simulation
1:12 Snow doesn't run off buildings and objects like that
They probably used the same friction value for everything in the scene and optimized it for the snowplow. The roofs should have a higher friction value
Lab is probably in California or something lol. That's how these things work
That is really cool how little memory it uses, but as an FX artist I do snow simulations and I can tell by those renderings that they're using like 10% of the amount of particles I normally use. Im just wondering if that's the extent of detail for that program or if that's just the extent if it being real-time? Either way really cool how efficient it is!
3:48 Ah yes, *L I T T L E* frosting
The cup just isn't big enough.
Wow! I am wondering how did they render so much particles!
I guess the big majority of the scenes is "static" snow without interaction, so they transformed it into a simple mesh and took care only about the snow close to the interactive part of the scene
This still looks like foam, not snow.
it is because of the rendering method and how they wanted to show snow
at least I think so...
Its because its not meant to its just showing the simulation of the particles in actual use for visuals the artist would add shading etc and change variables to make it actually look like real snow.
More 2 minute papers. My day is complete!
0:53 You need dial in some other settings. This simulation doesn't look convincing but it appears to be capable of producing a better result.
4:25 snow is not that rubbery. It almost does not come back after compression
Two Minute Paper: Half an hour per frame, yes that means all nighter simulations
RDR2: hold my beer
IMHO it looks often like foam and not snow. Especially the scene with the car and wiper.
Video: "shows snow"
UA-cam compression: "i'm about to end this man's whole career"
Very impressive especially for the time it takes to track all the particles, but hopefully they can continue to improve the accuracy and how realistic it is, because it acts kinda funny compared to snow, both soft and dense snow sims act more like sand or foam, and the snow seemed to lack momentum especially with the plow
I feel like the edges don't resemble snow at all. It looks more like slime.
I feel like this still has a ways to go for realism. The snow looks like it's acting more like a liquid than it should, especially the scene with the vehicle scooping the snow. Just something to look forward to though, can't wait to see the day.
at some points in the video it looks way to grainy
Everything looks like laundry soap powder!
there are many fine achievements with this latest work. i wonder if it would be very difficult to improve the snow plough sequence by leaving a layer of snow still on the road? it is like this in some places in the demo clip, but in others, the road is left clean and pristine, which is not very realistic.
it would be great if discolouration could be added to the snow, which soon takes on a grey colour (unfortunately), at least in cities. also, twigs and other debris fall on snow pretty quickly. not strictly speaking snow itself, but still a vital part of making it look realistic.
That might have more to do with the plow and the road both being too pristinely smooth and in perfect contact with each other. when it went over the bump, it did leave some, but i agree it looked a little foamy. Still, very impressive!
Discolouration is hard to accurately simulate, as it involes simulating dust and debris particles in the air interacting with the snow. (Atleast I think so.)
@@meeeshroom no question that the discolouration might be tricky, but for feature film work (which is what i need), things have to look highly believable otherwise the audience spots it and it erodes the dramatic suspension of disbelief. i am sure that with the right algorithms (and training) something highly realistic will be achieved.
@Bronze HD yes, that is a fair point. it would be interesting to see the results if a rough texture were used. if done with 3D geometry that could be expensive on polys, but i have not read the paper - maybe there is a neat and light way of doing it.
The Inuit of the Frozen North which is in Canada, have over 2000 words for snow.
Powder snow, wet snow, mushy snow, clumpy snow, frozen snow... you get the picture.
Can this snow simulation, simulate wet snow mixed in with lake effect snow?
Or should we wait for the next paper?
3:48 "Add a little frosting"
The machine: *day after taco bell impression*
Thanks for this video but I want to point out that the more complex scenes still run at 8-10min/frame. Which means that a 10sec scene would simulate for 40 hours. (See graph on page 9)
This is going to be difficult. The bunnies were way off, since the snow didn't melt and compress, it crumpled. Also, snow doesn't poor from roofs like that, it sticks to the roof and to the already present layers. The tearing is also weird or largely up scaled. You might as well do thermal capacity simulations because various things will take a long time before a snow layer starts appearing (e.g. on a wet tarmac).
yikes, still a lot of work to get the plow simulation working well - I live in a particularly cold climate, (outside it is currenty sub-zero F) and I can tell you the truck with the plow looked really unrealistic. Hopefully, that was older footage. The simulation essentially looks like viscus water, not snow.
Me literally every single time I watch another Two Minute Papers: "Why am I not subscribed to this channel yet?? Oh. I am."
I expect many more papers, because, while being impressive work, this looks and behaves nothing like snow. It looks more like what styrofoam balls would do.
I live in Sweden. The snow here does not act like this. Looking forward for the next paper...
Agreed in Michigan!
There are many types of snow my swedish friend, of course you know that better than me
2:01 sounds like a polite version of “Shit hits the fan” :)
The general shapes of the simmed snow look more like a wet putty or mud. The patterns don’t reflect the grainy, crystalline quality of snow or the way it crunches/breaks apart.
Hmm, it looks more like shaving cream than real snow, maybe couple papers more to have reasonably good looking snow.
It's still not perfect but definitely quite impressive.
I feel like the cohesion between the particles doesn't quite fit the behavior of real snow.
Super video as always, but for someone who live in the middle of a heap of snow at the moment, I have to say it does not look realistic at all. It resembles styroform particles.
That doesn't look like snow at all!
More like powder and sometimes like foam.
Snowmans don't jiggle
How is the "time per frame" calculated? When the results are compared with older papers do they take in consideration the hardware and software upgrades?
Doesn't look quite right.
Granted I was just out shoveling snow, so I have a vivid picture of the real thing.
Still the progress is impressive.
And still far away from real snow look, but maybe just good enough shader was missing there.
Hey, you are from 2035! Im from 2021, and i'll tell you. The snow simulation very legit bro
It looks more like flour or shaving cream. Or even a combination. Anyway its really interesting
I figured it out a few videos back, but I always thought you introduced yourself as "Dominic Pipers" or something at the start... I feel pretty dumb.
That guy walking in the snow is soooo unrealistic. He isn't struggling at all!
Snowrunner does it pretty well. Did consider it the best until seeing this..
Hold up. That blue guy seemed to walk unimpeded by snow. The snow reacted to him, but he did not react to the snow. As I've grown to expect perfection from this channel, this is simply unacceptable 😉
Why are these bunnies blue?
Cause if I was green I'd die. duh
(Me at Subway, ordering a sandwich)
Me: ... And a *little* bit of sauce please.
Subway employee: 3:46
hmmmm
I think I'll enjoy 20cm of real snow on Sunday, thank you very much 😉
Although this is impressive, it doesn't look really that great
I hope the improvements will have snow compression
Love these vidoes, even though I am not from this field its fascinating to see the advancements and wonder how much work went into achieving such beautiful results. As an aside: The gamer in me is dying to see how games of the future might apply these in real time!
I’ll soon do a tutorial with that tool on my channel! Pretty cool guys from Fifty2 🙂
This is beautiful and impressive. But what exactly is this good for? Will these simulations have a role in the real world other than entertainment?
Good question, Huzzzer! Yes, there are many real applications in the real world. Simulations are already being used to better demonstrate products to customers online. Allowing the customer to virtually hold and get a better look at the product can be improved. Simulating how clothes would look on the customer is making great strides.
Also, simulation are increasingly being used in industrial control systems and construction to make realistic views and sounds, and to realistically simulate operation. The industry is evermore embracing the use of actual video game technology to create the simulation because of the improvements in, physics simulation engine components, interface, and costs. These Two Minute Papers videos do an excellent job of showing what is providing some of these advances.
I'm sorry if this is a stupid question: how does the "time per frame" metric account for computational power? I'd imagine that within one paper, all of the measurements were made using the same computer hardware, but how would you compare between different papers from different times?
does anyone else think this looks.. not correct? Examples: At 2:05, a tire won't create styrofoam-like edges, with snow particles drizzling off - it will compact the snow edge, more resembling a knife cut. At 3:54, the ice cream wouldn't split into many pieces with the center staying rigid - it would all collect in a single blob on the ground, with maybe som small splatter.
At 2:41, only very wet snow would come close to retaining its rabbit shape. Very wet snow would not turn into fine powder, as seen when the ears fall off.
Maybe I'm overly critical, but as someone living in the arctic, I see that these simulations clearly deviates from reality to a large extent. So I find it a bit strange when Károly's commentary seem to imply, from the footage alone, that this paper is an amazing achievement. It might still be, for all I know, but I'd prefer if the commentary also acknowledged the weaknesses and limitations.
4:15 that guy is breaking snow apart like a frigging moose
3:47 Worst ice cream cone factory ever!
Just got 6 inches of snow last night, your timing is impeccable
Same here lol
"Elsa" wants to know your location!
Why did I read: ""Elsa" wants to snow your location!"
She must be in my head ahhhhhhhhhhhhhhh
I remember doing some explosive simulations on my old Intel 2600K back in the days. Simple simulations took over 8 hours, for only 10 seconds. Now we can do this with a CPU that is equal to an Ryzen 3800X: What a time to be alive!
I love it! In 5 years we will not be able to distinguish simulated from real.
As an incompetent person in simulations I can only tell that it looks like some kind of fluid or foam to me. It might be a huge milestone in simultaions as far as techniques are concerned but for visuals it is just not the best. Regards!
Nagy rajongója vagyok a csatornának és a munkásságának, így tovább!
Thats the rendering not the simulation.
Sounds like a ASMR video, no?
It's not meant to look like the material snow. What's shown here is a physical simulation of snow, not a visual one.
We can already make a surface look realistically snowy with the right properties, mainly subsurface scattering; the tricky part is the physics of it all.
This simulation looks like shaving cream because it's a like a technical paper, it's meant to make the simulation's details clearer rather than make the snow look realistic.
Ok, what's the story behind that turtle dog thing that shows up in a lot of these simulations?
It's an armadillo.
I don't know how it works exactly, but where do all these simulations end up? I mean, what is stopping to add all these features in unreal engine (for example) right now? Selling that to epic games would create a new generation of videogames. Photorealistic graphic is mostly already achieved for triple A games, but what is lacking is proper physics simulation of nature. Also 3d artists would love this.
I'm halfbrain Artist that can do only simple simulations. I don't understand why it's so good except effeciency. To me it's looks like a slightly frozen foam with slushy cream inside. Does it supposed to have second pass simulation with small particles of snow in cracked areas or something? I understand that it's not production shader for final render. I just complain about lack of my imagination "how it might work in real world VFX".
1:21 Anyone from a cold climate judging this as harshly as I am? Cause I know from experience that snow would never clear THAT perfectly...
I love this channel so much
I was wondering if UA-cam Company could ever develop a client side algorithm, that buffers frames of the video in low resolution, then enhances resolution using ML to result in a genius predicted resolution.
This way network will suffer less, the computation is on the client side, get nice no glitching video frames with high resolution.
Note: this suggestion if for countries with poor network connection.
Any humble comments are accepted!!
The old one looked more convincing
That is quite the jump.... Does it scale ? We all want 30FPS but - I would love love love to see this cracked at maxed and see it's BEST output in SPF
Given that half the comments seem to be saying this is the most realistic simulation they've ever seen, and half are saying that no... this doesn't look right, too sticky/foamy/liquid looking.... I'm pretty sure we must have two very different types of snow in the world, and I've apparently only lived in places that get the kind that doesn't look like this.
I think this one stills has a ways to go before it looks remotely realistic. Two more papers?
Greetings from Finland. It seems I managed to procrastinate clearing up the snow at the yard till tomorrow, watching Two Minute Papers simulating snow. ;) I have to say these simulations look more like 'wet dry powder' than actual snow in behavior. Material in video is absolutely too fluid and uniform in density, droops right after it's fallen, compacting and solidifying is missing, melting simulation misses the state where snow still holds its shape but wets (I guess it needs more thermal than particle simulation) and so on.
Károly, you need to share your course on light transport simulations with the authors of this paper. Not trying to diminish the authors' achievements in snow simulation but the material choices are just poor. As others mentioned in the comment section, the snow looks more like a kind of shaving foam. On the other hand, Disney's paper uses very believable materials and that makes it harder to compare the papers directly. Can't wait for new video games in snowy environments (few papers down the line, as always).
Looks good, however it seems there's too much focus on the interaction of small individual particles. Snow really clumps when it starts and goes above 0. However this of course depends on the conditions before hand. I'm Canadian heh
It's definitely a leap forward but it's amazing how far it still is from realism. Just shows how much work there is still left for the people joining the field each day.
You can see that they don't have the level of dynamic change and variation that snow has.
The powder snow was too sticky and the densely compressed snow wasn't sticky enough. They need to add a layer that adjusts the friction and cohesion dependent on compression.
Powder snow won't make a good snowball but if you run it over with a car it'll compress very solidly. The snow pushed by the snowplow should turn into hard dense packed chunks as it pushes out.
I'm also curious how they deal with partial melt because a big element to how snow acts is partial melts at different stages and interaction with small amounts of melt changing clumping properties.
240Mb??!!😯 Can't wait for synthetic, molecularly-folded milkshakes!
them : millions of particles but 200 Megabytes.
me: a few pillars are procedurally created in unity and 32 GB memory requirement.
They react like Styrofoam particles and nearly all are copies in that the single individual object in that all react similarly, they "melt" the same way. You do not take into account the cohesion of the ice crystals and instead treat each particle as it's own entity.
Not even powder snow reacts this way. Your presentation looks great btw, sadly, snow does not react like this. At the very least you need to add in some cohesion factor based on temp, to include the ice crystals melting singly or in clumps depending on temp and environmental factors and objects like the car used in this demo. The car was painful to watch as snow does not react that way when driven over. You need to add many new factors and variables. Cohesion due to temp (single entity crystals landing when frozen, melting together slightly, or simply interlocking when cold, their overall weight and reaction when combined etc) and compression ( what happens to that square of snow due to its own weight? Weight with temp gradients, refreezing?) are two that need to be added.
You could have told many thing, but this being snow? No.
It strikes me rather like some magic styrofoam that can turn into a sticky gel.
it has some strange isotropy to the way it is moving and breaking, the harderning seems like a "oh, stress, lets just randomly combine with the surrounding", and when in contact with other surfaces it turns into a gel. 2:08 shows that really clearly.
Just look at the behaviour at 1:30 - yes, it leaves some sort of (strange) tire marks but everything else does not make it seem like snow.
By how it looks i kinda also doubt that it can even come near disneys version in terms of behaviour - just snow rolling down a slope and accumulating.
I live in Russia, and from my own experience I can tell that this doesn‘t look like snow at all. More like sour cream or something. Especially on the snow city and car demos. :(
Disney’s snow looks a lot better.
My goal in life is to have two minute papers make a video featuring me.
Usually I'm blow away by the simulations presented on this channel, but this was actually a bit disappointing? Looks like it's 20 years old? :)
20 years later: Dear fellow scholars, today we have great news: The machine learning algorithm can now simulate all possible interactions of quantum particles from scratch in only 4 minutes!
I was expecting to need 4 sticks of 16 gig ram, but even my mum's laptop can run this simulation. Outstanding!
I'm sorry. I live in a country being snowy right now and I can tell you: This doesn't look like snow. More like packaging material or little balls of styropor.