Crysis 2 Screenshots, U-Decide 2010 Preliminary Findings

[PressBot]

We have a double treat for you today.  First, here are some exclusive screenshots we got at GDC Online from Crytek's upcoming Crysis 2 in 3D.






All these images were captured with a stereoscopic 3D camera on a JVC HDTV, so if they are a bit blury, it couldn't be helped.  This is a "pre-alpha" version of Crysis 2, so it does not necessarily reflect what the full version will look like.
Read full article...

[Dom]

I looked at these in 3d and look ok for what is out and all. I just hope crytek uses full scene depth FOV. Sadly most 3D games are not rendered at full scene depth.

[Fredz]

What do you mean by full scene depth FOV ?

[Dom]

What do you mean by full scene depth FOV ?

I mean that only close rendered objects like the gun, hands, or any object that is only within about 2 meters-ingame is using depth. All the backgrounds stuff like windows in a high rise or distant structures are 2d looking.

"Full scene depth, field of view" would make the whole enviroment look like its got depth even the windows in a sky rise thats far away. Sure would make a 3d game alot better.

[cybereality]

To be honest, I am not seeing much 3D effect at all on those pictures. I will still get the game regardless, but so far I am not impressed.

[tritosine5G]

Well, all I can say is that it opened *that* extra dimension quite well with DLP projector .

Maybe it boils down to separation vs. screen size, the thing Blackshark mentioned about Avatar vs. monitors.

[Neil]

It's not a fair analysis of what Crysis 2 will deliver, and I only saw some bits and pieces. Keep in mind that this was taken through a 3D camera!

Regards,
Neil

[Likay]

I'm really curious on how it looks and the pictures, while giving a preview unfortunately doesn't say too much.
Did you see it yourself, and what is your opinion about their render algorithm? Honestly i'm dying to see how a computergenerated 2d+depth looks like.

[Fredz]

Are you sure it's 2D + depth ? In the CryEngine 3 presentation published at HPG 2010 there are three modes listed for stereo rendering. And the first one in the list is brute-force, the same mode that's been used by the NVIDIA stereo driver.

[Likay]

I missed this. If you're right it's great info! There was another article about cryengine and how they managed to get a 3d image almost without any fps loss. I assumed this was their native support for 3d.

[Zerofool]

I personally am not impressed at all by the shots, but that was expected. Plus, it definitely was set with low separation, so 3D-noobs could enjoy the demo... We'll have to wait and test it personally .

Fredz, that document you're talking about is somewhat old as far as I know, and the latest info (form that 3-part interview with MTBS3D) stated that they will only use their own method (the 2D+Depth everybody is talking about) for all platforms, including PC! Which is a shame.

Likay, you are wondering how it will look like? If someone of you have ever used the modelling/rendering software Blender, you will probably remember a neat thing called Vector Motion Blur. What it does is just taking all the speed vector data for all moving objects in the scene (including the camera), and apply simple blur in post processing, according to the speed of each object. The principle the guys at crytek are using is probably similar in concept - using the depth data of the scene, they create the second view with simple post processing (just distorting the image). And this will probably give the second image a bit of softer, photoshopped look, but with separation that low, I think it will go unnoticed. Using high separation with that method is just out of the question I guess, and we'll see some very limited "depth" option in the game. But I hope I'm wrong and will be pleasantly surprised .

[cybereality]

You can check out this video of the original Crysis using the TriDef "Virtual 3D" mode (2D+Depth):
http://www.youtube.com/watch?v=3auOYN9EFQI" onclick="window.open(this.href);return false;

I think it looks decent and is certainly playable. Not as good as true dual-render, but acceptable quality.

[Fredz]

Fredz, that document you're talking about is somewhat old as far as I know, and the latest info (form that 3-part interview with MTBS3D) stated that they will only use their own method (the 2D+Depth everybody is talking about) for all platforms, including PC! Which is a shame.

The publication I'm refering to is from June 25, 2010 so it's not old at all. And it's been written by the developers themselves, not by a designer like the one interviewed by MTBS3D. I've already read these interviews and there is no mention that they're going to use this technique on all platforms.

CryTek has published two new papers about their stereo rendering technique on August 24, 2010. In one of them they say that rendering twice on consoles is not an option for them because of the performance penalty, so they use screen space reprojection in this case. But they also say that this technique produces artefacts, so I'm inclined to think that they won't use this technique when rendering performance is not a problem, ie. on the PC platform.

http://www.crytek.com/cryengine/presentations" onclick="window.open(this.href);return false;

Likay, you are wondering how it will look like? If someone of you have ever used the modelling/rendering software Blender, you will probably remember a neat thing called Vector Motion Blur. What it does is just taking all the speed vector data for all moving objects in the scene (including the camera), and apply simple blur in post processing, according to the speed of each object. The principle the guys at crytek are using is probably similar in concept - using the depth data of the scene, they create the second view with simple post processing (just distorting the image). And this will probably give the second image a bit of softer, photoshopped look, but with separation that low, I think it will go unnoticed. Using high separation with that method is just out of the question I guess, and we'll see some very limited "depth" option in the game. But I hope I'm wrong and will be pleasantly surprised .

Yes, they use the depth buffer to generate the second view, but there is no reason the second view would be softer, have a photoshoped look or be limited to low separation values. Read the paper entitled "Part 2 - Bringing Stereo to Consoles.ppt" on CryTek site to understand how they do it.

[Likay]

@cybereality: For being a virtual 3d-mode crysis looks insanely good! Imo way good enough and playable compared to gaming through a stereodriver where you might have to put up with one and another anomaly instead. True, it doesn't beat true stereoscopy but is very close and definitely enjoyable.
And some great info you digged up Fredz!
2d+depth probably gives best results when done in the gameengine compared to when depth is added manually in post.
As said i'm really curious on how the final result is going to look like. Anyway it's good to have an option for true stereoscpopy as well.

[Zerofool]

You can check out this video of the original Crysis using the TriDef "Virtual 3D" mode (2D+Depth):
http://www.youtube.com/watch?v=3auOYN9EFQI" onclick="window.open(this.href);return false;

I think it looks decent and is certainly playable. Not as good as true dual-render, but acceptable quality.

Yes, it looks nice, and very playable, but the "softness" of the image I'm talking about could only be perceived when watching (visually) lossless video, not washed out, crappyly encoded low-bitrate video (that's why I hate YouTube, until they change that).
The separation I'm talking about is waaay higher than what is visible here (and probably that's close to the maximum of the DDD driver, I tried it very long time ago and can't really remember ).

CryTek has published two new papers about their stereo rendering technique on August 24, 2010. In one of them they say that rendering twice on consoles is not an option for them because of the performance penalty, so they use screen space reprojection in this case. But they also say that this technique produces artefacts, so I'm inclined to think that they won't use this technique when rendering performance is not a problem, ie. on the PC platform.

http://www.crytek.com/cryengine/presentations" onclick="window.open(this.href);return false;

OK, I took a look at the presentation, nothing new here. As you said - they consider the "render twice" technique as no option for them. I really doubt they will sit down and write a second version of the S3D rendering portion of their engine only for the PC version (considering they'll win much more money from the console versions). Still, I hope you are right, although not very likely.

Yes, they use the depth buffer to generate the second view, but there is no reason the second view would be softer, have a photoshoped look or be limited to low separation values. Read the paper entitled "Part 2 - Bringing Stereo to Consoles.ppt" on CryTek site to understand how they do it.

I personally have very low tolerance for "graphical impurities" and that's why I call it "photoshopped look". Imagine you captured (lossless) screenshot form a 2D game (without using anti-aliasing). OK, now apply some stretching, smudging and resizing - you can't preserve this edgy, sharp look of the original image whatever you do - it will look somewhat blurry in these processed areas. That's what I meant by "softer look". Using AA however, would hide much of this.

About the separation values... now just imagine playing Borderlands, switching to iron sight, and having set huge amount of separation (probably at unplayable levels). On the left eye image you will only see the left side of the gun (as well as it's top side, and the rest of the scene of course), and on the right eye image you will only see the right side of the gun (and again - the top side and the rest of the scene). Now, there's no way this algorithm will correctly produce the look of the gun in the second image, because there's no pixel/color data about it's other side in the first image. That's why the separation level is quite limited (but probably sufficient for most of the gamers).
Take a look at that crytek presentation, page 5. "Works surprisingly well with sensible stereo strength" I don't want sensible strength, I need extreme strength .

Anyway, let's just wait and see how the game will look closer to its release, hopefully in high quality 3D images/video .

[Fredz]

I really doubt they will sit down and write a second version of the S3D rendering portion of their engine only for the PC version (considering they'll win much more money from the console versions). Still, I hope you are right, although not very likely.

In their June paper they say the engine can already support these rendering modes :

• - Brute-force stereo rendering
  - Central eye frame with reprojection
  - Experimental stochastic rendering from one of eyes

So it should be quite simple for them to offer an option to enable one of these modes in a game.

I personally have very low tolerance for "graphical impurities" and that's why I call it "photoshopped look". Imagine you captured (lossless) screenshot form a 2D game (without using anti-aliasing). OK, now apply some stretching, smudging and resizing - you can't preserve this edgy, sharp look of the original image whatever you do - it will look somewhat blurry in these processed areas. That's what I meant by "softer look". Using AA however, would hide much of this.

A form of AA is already used for rendering triangles in classic 3D (mipmap filtering), so I don't see why it would be worse in stereo 3D. It all depends on the ratio between the size of the texture and the size of the triangle on which it is rendered. The distortion created by the rendering of a second view should be very marginal since the scale is almost identical. It'll be problematic for triangles that are not seen by both eyes, but for general rendering there's no reason it would give blurry textures.

On the left eye image you will only see the left side of the gun (as well as it's top side, and the rest of the scene of course), and on the right eye image you will only see the right side of the gun (and again - the top side and the rest of the scene). Now, there's no way this algorithm will correctly produce the look of the gun in the second image, because there's no pixel/color data about it's other side in the first image. That's why the separation level is quite limited (but probably sufficient for most of the gamers).

Yes, that's one problem with this technique that I thought they would have found a new and better answer to. By reading their paper it seems they didn't do things differently than DDD or Sony, which is a pity considering the artificial buzz they created for their stereo technology in the press.

Still, this problem should be quite marginal since the difference between two viewpoints for a close object (a gun) is quite minimal, some pixels at most for maximal stereo separation. But I agree with you that's it's still going to be an annoying artefact.

Considering they have complete control of the engine I was expected a much better solution from them, for example by detecting non visible faces for one eye and rendering them normally to avoid such artifacts. I've read several research papers about virtual reality applications that did use geometry shaders to not have to recalculate completely the second view for rendering. Too bad they didn't try this route for triangles that are only seen for one eye...

[Likay]

It all remains to be seen.
If crytek doesn't manage to implement 2d+depth in a convincing matter i don't think anyone can (since it's done already in the graphic engine which should be the best way of doing it). In this case we could trash 2d+depthmap as an allover conversion method. Some converted photos looks great in 3d though if enough work is put into it. So far the same can't be said about movies.....
Hurry crytek! I'm curious!!!

[tritosine5G]

Well, to tell the truth, the 50% hit brute force rendering was always somewhat alien to me. Especially in realtime graphics, where everything is centered around clever tricks, and I love to learn about multiplexed concepts in general ( color sequential , etc.).

For example I can even imagine something like , dithering (intentional added noise!) to seperate the 2 viewpoints, graphically. Whew that sounds crude written down, but still, thats what I think.

In my mind , a fully stereo compatible engine would be a software renderer, built with stereo3d in mind from ground up, there wouldnt be 50% hit there either (hopefully?)

[Fredz]

I'm not sure Crytek is the most qualified team to solve this kind of problems. Their core competence is 3D engine programming and they only had a quite recent exposure to stereo rendering. The fact that they created some buzz on virtually nothing conduct me to think that they don't know the field enough and don't have the required experience to produce something valuable. Moreover this technology is not a strong selling point for them at this time, considering the state of the market.

I expect more from specialized companies like DDD, iZ3D, TriOviz or even Sony since they directly or indirectly depend on this technology for their sales. And there is still some place for newcomers that are specialized in stereo rendering but quite new in the gaming field.

Concerning 3D conversion, it's badly done in live action movies because the technology is very different from the one in video games. Still there is a good example of a type of conversion done for movies that is comparable to what we can find in video games, namely the one used for "Meet the Robinsons".

In this case, they had access to the 2D rendering and the 3D information, just like in a game. They created a second view using only these two informations since a new rendering for the second view would have been too costly. I didn't see the result yet, but according to the realorfaked3d website the result is convincing enough to make them classify this film as native stereo. So I still have some hopes that this can be done in an efficient manner and with acceptable results in video games.

[Zerofool]

In their June paper they say the engine can already support these rendering modes :

• - Brute-force stereo rendering
  - Central eye frame with reprojection
  - Experimental stochastic rendering from one of eyes

So it should be quite simple for them to offer an option to enable one of these modes in a game.

Well, there's still some hope then . Good to know.

A form of AA is already used for rendering triangles in classic 3D (mipmap filtering), so I don't see why it would be worse in stereo 3D. It all depends on the ratio between the size of the texture and the size of the triangle on which it is rendered. The distortion created by the rendering of a second view should be very marginal since the scale is almost identical. It'll be problematic for triangles that are not seen by both eyes, but for general rendering there's no reason it would give blurry textures.

You have misunderstood my statement. I'm not referring to blurry textures and texture filtering. The problem is that the second view is not rendered at all! As you know, all it basically is, is image, interpolated/resampled (bilinear interpolation used in the case of Crysis 2) from the first view, nothing more. That's why it's fast - it's simply a 2D image processing. This interpolation makes the (whole) second image softer. If the first image is not anti-aliased, the trails of the interpolation in the second image are much more visible (giving it more "photoshopped look" ), compared to anti-aliased image that's being interpolated. Also, in the second image, the parts with bigger difference between the two views will look softer and blurrier that these areas with much lower difference.

Considering they have complete control of the engine I was expected a much better solution from them, for example by detecting non visible faces for one eye and rendering them normally to avoid such artifacts. I've read several research papers about virtual reality applications that did use geometry shaders to not have to recalculate completely the second view for rendering. Too bad they didn't try this route for triangles that are only seen for one eye...

Yes, that's much better solution, but it would cost them much more than 1.5% of the performance (of course, depending on the scene).
Anyways... considering the power of the GPUs of today, the good old Brute-Force way is not a problem I guess. Well, at least until the next generation of consoles raise the bar of graphics quality in these next-gen games (~2-3 years from now).

P.S.: Man, I've been writing this post more than an hour, I just saw your new post .

[cybereality]

I have faith that they will do a good job, or at least the best job that can be done with 2D+depth. It remains to be seen if this will match true stereoscopic rendering.

[Fredz]

You have misunderstood my statement. I'm not referring to blurry textures and texture filtering. The problem is that the second view is not rendered at all! As you know, all it basically is, is image, interpolated/resampled (bilinear interpolation used in the case of Crysis 2) from the first view, nothing more. That's why it's fast - it's simply a 2D image processing. This interpolation makes the (whole) second image softer.

The first image is already "soft" because of the bilinear filtering from the textures depending on the viewpoint. The second image shouldn't be that much "softer" since the viewing angles are very similar in most cases. It should only be noticeable when the horizontal apparent size is very different in the two images, but that's the case only for triangles viewed with a very narrow angle, which generally don't have a big horizontal apparent size.

Consider the following figure, using a 24" display at 1920x1080 using a 256x256 texture placed on the black line :



In the first case, the texture is rendered on the screen plane with 156 pixels for the left eye and 234 pixels for the right eye. The first image is filtered bilinearly from the texture (left eye), the second one is filtered bilinearly from the first image (right eye with reprojection) and the third one is filtered bilinearly from the texture (right eye with brute-force).



There is a somewhat noticeable loss of quality in the second image compared to the third one, but you can also see that the quality of the first image is not perfect either. And it's really a worse-case scenario visible mostly on edges, because in the second case depicted in the schema the horizontal sizes are very similar for each image (588 and 582 pixels resp.). So you shouldn't be able to notice much difference between reprojection and brute-force most of the time.

I think they could even improve on this technique by using directly the texture to calculate the reprojection instead of the rendered image. It shouldn't be very different since the texture is already in video memory, and then the second image would look exactly like the third one.

Anyways... considering the power of the GPUs of today, the good old Brute-Force way is not a problem I guess. Well, at least until the next generation of consoles raise the bar of graphics quality in these next-gen games (~2-3 years from now).

Sure, with current GPUs you can always use brute-force on old PC games with full rendering quality in stereo. But on recent games or on consoles you'll always have to compromise on visual quality by disabling some visual effects or by rendering at lower resolutions. Are you really ready to buy a $200 graphics card for your PC that can only offer the quality of a $100 card when you play recent games in stereo ? Me not !

I'd really prefer game developers to find another solution than brute-force to have a better compromise on visual quality. Screen reprojection may not be the solution yet, but I've good faith they'll come with something better when stereo gains in popularity.

P.S.: Man, I've been writing this post more than an hour, I just saw your new post .

It happens to me a lot too, I've got a bad habit of writing long messages and it takes time to verify that you are writing as good english as possible when it's not your mother tongue...

[Likay]

To clarify a bit (and Fredz explanied very well): I did mean that crytek is able to make the best of 2d+depth just because they do have access to depthinformation already within the gameengine. This is why i think that if this doesn't look good, then nothing using 2d+depth will.

[Dom]

I heard the beta of crysis 2 is out on xbox fully and somewhat lingering around for the pc. Can anyone confirm this so we can test it out?

I found alot of columns talking about this after checking
http://www.techspot.com/news/40667-crys ... -only.html" onclick="window.open(this.href);return false;

Hopefully someone can shed some light on how the xbox performs in stereo3d with this beta, if its possible to enable stereo3d at all.

[Fredz]

Hopefully someone can shed some light on how the xbox performs in stereo3d with this beta, if its possible to enable stereo3d at all.

If the option is present in the beta for the Xbox 360, you will need a display that can handle frame-compatible formats (side-by-side, line interlaced, etc.) since it doesn't support native HDMI 1.4 stereo. In the CryTek publication I mentionned earlier it's also said that the dashboard will look broken in stereo mode on the 360.

[tritosine5G]

uh ppl started to show x360 recordings on youtube , crazy dope man...!

[Neil]

The following is an excerpt of a blog article.  Read Full Article

We have a double treat for you today.  First, here are some exclusive screenshots we got at GDC Online from Crytek's upcoming Crysis 2 in 3D.

All these images were captured with a stereoscopic 3D camera on a JVC HDTV, so if they are a bit blury, it couldn't be helped.  This is a “pre-alpha” version of Crysis 2, so it does not necessarily reflect what the full version will look like.