These Are Some Good-Looking Computer Rocks

Megalith

24-bit/48kHz
Joined
Aug 20, 2006
Messages
13,000
I normally don’t post videos of rocks, but this one is pretty special: while the scenes here appear to be photographed or filmed, everything is actually, in fact, computer generated. Graphic artist Rense de Boer, who worked as a technical art director at Battlefield developer DICE for several years, used two GeForce 1080 Ti cards to explore the boundaries of real-time rendering.

It took him five days to photograph and capture the landscapes he wanted to render, and another several weeks to rebuild them in video game making tool Unreal Engine with a process called photogrammetry, which works by compiling a host of images taken from various angles and using them to create realistic depth. "Plants and trees are built leaf by leaf, giving it depth that is needed to ensure realism and create a more believable world," de Boer said.
 
ahahah no doubt

hopefully we will see video games rendered 100% like this in our lifetimes.
 
ahahah no doubt

hopefully we will see video games rendered 100% like this in our lifetimes.

I'd say it'll be close to that within not too far off. So unless you're 70 years old, you'll live to see it.

Right now, it's rendered in real-time with two very powerful GPU's. That power needs a few generations to filter down to the cheaper end cards so that artists will put that level of detail in to their games. Other systems/toolsets also need to catch up so that optimizations can be done and a game with multiple assets (not just rocks) can be made.

Look at some of the ArchViz / real-time artists for UE4 and you'll see what they're capable of in a game engine.
 
It's not just the lighting, this has no perceivable aliasing that I can tell either. The video is at 4k, but looking carefully, I can't see it at all, makes me wonder what the raw settings are.
 
That looked good.

Didn't The Vanishing of Ethan Carter use the same technique?
 
Don't worry, if that detail makes it into a game, it will only support 1 1080ti ;)
 
Good ol' Unreal Engine.

Has some of the best PBR in the industry.
 
Says he had a short stint at NASA, interesting, but seemingly unrelated. Not sure what their interest with CGI rocks would be.....
 
a looong way from a convincing coke can. Still remember when I saw the first one and couldnt believe that it was computer generated. Another reason to go best graphics card if u can afford to.
 
Forgive me if I am misunderstanding something here but doesn't the photogrammetry technique involve tons of photographing and stitching them together to make a 3D model? Wouldn't it be more accurate to say that everything had been photographed and filmed over 5 days then recreated in engine?

Just looking up the definition of photogrammetry gave me this: "the use of photography in surveying and mapping to measure distances between objects." If I understand correctly, the artist basically took a bunch of photos at different angles that allowed him to map the objects and create 3D models of them that he could then overlay with the photos for a photorealistic rendered model. The original article uses the term "made on a computer" rather than computer generated. I think we are still a bit away from being able to generate photorealistic renders without needing photography.
 
Photogrammetry really is the shit.

It's not just the lighting, this has no perceivable aliasing that I can tell either. The video is at 4k, but looking carefully, I can't see it at all, makes me wonder what the raw settings are.

The possible issue is that when someone is using photographs of a certain resolution, wouldnt it be easy to "cheat" a lil using the photographs instead of raw rendering? How to tell the difference if a photograph of a real scene was used to create the render? I am sure that the photograph and render was on screen at the same time - cut and paste. Would like to know where the differences show up, or is it just turning a flat image into a 3d one?

Looks really really good if it was done by a game engine. Possibilities.
 
The possible issue is that when someone is using photographs of a certain resolution, wouldnt it be easy to "cheat" a lil using the photographs instead of raw rendering? How to tell the difference if a photograph of a real scene was used to create the render? I am sure that the photograph and render was on screen at the same time - cut and paste. Would like to know where the differences show up, or is it just turning a flat image into a 3d one?

Looks really really good if it was done by a game engine. Possibilities.
Well I admit, I'm only on 1440p here, but downloading it at 4k, I still can't see any AA whatsoever. I say this as someone who is particularly sensitive to the shimmering from FXAA and SMAA. If I had to guess, I'd say they did the rendering at a MINIMUM of 8x MSAA at 4k resolution, but possibly 16x or else some crazy high downsampling.
 
Well I admit, I'm only on 1440p here, but downloading it at 4k, I still can't see any AA whatsoever. I say this as someone who is particularly sensitive to the shimmering from FXAA and SMAA. If I had to guess, I'd say they did the rendering at a MINIMUM of 8x MSAA at 4k resolution, but possibly 16x or else some crazy high downsampling.
You don't see aliasing because aliasing is most noticeable on straight long edges of polygons. And using this method the longest edges of polygons are a few pixels long only, unless you go for extreme closeups.
 
You don't see aliasing because aliasing is most noticeable on straight long edges of polygons. And using this method the longest edges of polygons are a few pixels long only, unless you go for extreme closeups.
FXAA and SMAA both practically eliminate aliasing on long straight edges, and have for some time. However, they exacerbate it on small corners where they meet, or lots of overlap of alpha textures, or shimmering from light reflection off shaders, that sort of thing. That stuff always sticks out like a sore thumb to me. I can distinguish the aliasing using shader-based methods on a 4k image no problem. That's why I'm saying I see NONE of it on this footage, even on the close-up. Usually I have to get to 8x or 16x TrSSAA / Adaptive AA or SSAA to see NONE like this.

I did see other tiny flaws, like the imperfect depth of field effects, and some slight filter noise on the left side around 1:38, but I can't see any aliasing whatsoever. Whatever they're using is a very high level.

You say "using this method", what method are you referring to?
 
FXAA and SMAA both practically eliminate aliasing on long straight edges, and have for some time. However, they exacerbate it on small corners where they meet, or lots of overlap of alpha textures, or shimmering from light reflection off shaders, that sort of thing. That stuff always sticks out like a sore thumb to me. I can distinguish the aliasing using shader-based methods on a 4k image no problem. That's why I'm saying I see NONE of it on this footage, even on the close-up. Usually I have to get to 8x or 16x TrSSAA / Adaptive AA or SSAA to see NONE like this.

I did see other tiny flaws, like the imperfect depth of field effects, and some slight filter noise on the left side around 1:38, but I can't see any aliasing whatsoever. Whatever they're using is a very high level.
The blurriness of the dof effect covers up many things. Also the edges here are no more than 5-10 pixels by my estimate. So they might not use AA at all. Maybe just rendered at 8K and downsampled
 
Something about the larger structure at about 0:45 that just didn't look right... a bit too "sandy". But damnit that looked good.

I expect something like this to be an E3 demo for a game then they bring it down about 800 notches for the actual release game "oh we couldn't get that done..."
 
Great article about photogrammetry. Also Ethan Carter is a good cheap game that is worth a buy :)

http://www.theastronauts.com/2014/03/visual-revolution-vanishing-ethan-carter/

Thanks for the link (y)

It seems more like an effective way to get the 3d models into the engine/computer than raw compiling. A shortcut. Amazing in what can be done once its in the machine, but along the same lines as using humans to animate a "actor" instead of scripting the same actor. Wearing funny helmets with lights illuminating every facial motion instead of hand animation. Photogrammetry is a better way of getting the data into the machine.

:)Far better than a artist and a wacom pad.
 
You say "using this method", what method are you referring to?
Referring to photogrammetry. Or photoscanning as the ethan article calls it. They even go so far to outright call it scanning which makes me cringe. Because it has nothing to do with scanning. They probably call it that incorrectly because the best commercial tool on the market currently is called "Photoscan". Scanning is done with lidar, photogrammetry is an entirely different and less reliable method of attaining geometry data. But for creating photo realistic 3D models it is perfect and has been used in VFX for a while now.
 
Forgive me if I am misunderstanding something here but doesn't the photogrammetry technique involve tons of photographing and stitching them together to make a 3D model? Wouldn't it be more accurate to say that everything had been photographed and filmed over 5 days then recreated in engine?

Just looking up the definition of photogrammetry gave me this: "the use of photography in surveying and mapping to measure distances between objects." If I understand correctly, the artist basically took a bunch of photos at different angles that allowed him to map the objects and create 3D models of them that he could then overlay with the photos for a photorealistic rendered model. The original article uses the term "made on a computer" rather than computer generated. I think we are still a bit away from being able to generate photorealistic renders without needing photography.
That photogrammetry is not the same. The principles used are the same. Photogrammetry in mapping has been around for at least half a century. I'm too lazy to look up how long exactly. This is using the same method of using photos taken of the same object from slightly different angles to re-create 3D data. The difference is that software has become so advanced that if you do a good job on the field (photographing the subject) then the software will spit out the finished model almost without user input. It takes substantial CPU time, say 24 hours maybe for a scene with a few dozen images. But a modeler could never achieve the same level of detail even in 24 days.
 
Good!!! Now we can fake some rover landings on exoplanets!

On a more serious note, that does look really good.
 
Back
Top