Accurate human rendering in game [2014-2016]

Status
Not open for further replies.
Star Citizen MMO character face - could be one of the player faces
Tier1_Face by X-RAY-89, auf Flickr

This should be a Tier 1 head which has not the render budget like a Tier 0 head (main characters like Gary Oldman etc.)

Few head animation tests start at 13:28:
 
Last edited:
Yeah, we need to get hair and fuzz as part of moving into the direction of 'accuracy'. ;) That man is super creepy.
 
It is a Tier 2 and not a Tier 1 head

Developer on reddit:
"Its actually a tier 2 ;) His name is "male03"

There isn't a drastic difference between t1 and t2 to be fair (their used to be a year ago but with some SUPERB performance tricks/saves they have become very similar).

T0 is even higher poly than what you see there ;) more animated diffuse and wrinkle maps too."

This means that passersby on the street and ordinary NPCs will have the quality like above. CIG has already scanned 120 heads (Jump Point magazine is the source).

"I think its not helpful for people to see floating heads ;) The biggest thing that will pull you out of it is the lack of body motion. We don't see people often PERFECTLY still and just moving areas on the face all isolated. Also keep in mind that animation on him is really just something that was quickly made to show you guys some new head, and not a performance capture. The performance capture + body animation helps immensly :)"

Source: https://www.reddit.com/user/scumbag_seantracy
 
Last edited:
The Star Citizen character models are absolutely jaw dropping. Does anyone know their poly count and how it compares to the other big AAA games ?

I love the FFXV character design although the World seems a bit bland in comparison to games like Horizon, The Witcher 3 and Breath of the Wild. I wonder how much of the games image quality / aliasing issues will be fixed just by running it at 1440p upscaled to 4k on PS4 Pro compared to 900p on the standard PS4 ?
 
The heads have multiple versions of quality. Tier 0 has the highest budget and Tier 3 the lowest. Tier 0 will be used for the main characters in Squadron 42, the singleplayer campaign.
Marine suits etc. are not tiered, same with clothing. Tier's are exclusively for heads.
A Tier 0 head has without hair has 40 000 polygons and a Tier 1 head has about 12 000 polys. Marius' head in Ryse had (as far as I know) 10 000 polygons.


Something interesting to read.
Kyran: With a large number of incredibly detailed characters at once, both NPC and player, what’s the design plan to keep this from blowing up client resources within the persistent world game?

Sean T: I’m glad you asked!! This is very important question and has been considered since the start.

The technology plan has already progressed quite far along, as we can already support the content we’ve made. But we are about to make one last big push which will take us through to final.

First off, let’s start with how expensive one of these heads was at this time last year.

A Tier 0 (Bishop, for example):
The mesh was 40K polys over 440 blendshapes! We then had to add the hair and scalp, and to get proper deformation needed to add a number of blendshapes to the hair too. This bloated the mesh to over 300 MB, not even counting the cost of the LODS that would have to exist additionally.

Since we only have so much video memory to use, we would never be able to have more than one or two faces in a scene at one time, doing it the classic way. We’d be lucky to have one.

Second was the texture cost. Since we do 44 different areas of blended wrinkles AND blended diffuse, our texture cost was quite high as well. 96MB to be precise, in blend textures only! Add the base diffuse, normals, spec, transmittance and scattering mask and we were nearing 150MB of texture memory per head.

To give some context to the numbers, we almost took up with one head as much video memory as was budgeted for an entire scene in Crysis!


So, the first thing to do was fix up the mesh cost.

Our graphics team came up with a great method to compact the cost of the corrective blendshapes so that it was reduced from the 300MB mesh to just over 30 Megs. This was a GIGANTIC win — anything that saves 90% of the memory with zero change in fidelity is something to celebrate.

It also now meant that we could support as many blendshapes as we wanted without a major impact on the mesh size.

Having fixed the mesh cost, another win came out of the graphics team. They came up with a very clever solution to reducing the texture memory with what I consider negligible quality impact. They came up with a system that stores the base map’s high frequency detail at high res and merges the base and blends at a reduced resolution. This took the cost of the blend textures from 72Mb to 6!!

This originally was one thing that was very different between tiers, as a Tier 0 would have "triple wrinkle mapping" or 44 areas with both normals and diffuse blends; a Tier 1 would have 28 areas "double wrinkle mapping" with just normal blends; and a Tier 2 or 3 would have "single wrinkle mapping" with just 8 areas.

Since this gigantic save (and even more coming), we’ve upgraded the T1s to include the animated diffuse, as we all feel it adds so much and the performance cost is far reduced from what it was originally.

Finally let’s talk animation data. Having a rig of over 500 joints causes animation data to be quite large. Storing animations for each of the heads quickly gets to be a massive amount of data.

Our R&D animation programmer, in concert with 3lateral, has now implemented the five heads with the run-time rigging technology enabled. This allows us to share ANY facial animation between any faces and still retain full support for wrinkle, blendshapes, tangent updates and all. It even exposes more powerful tools, like more sophisticated look-at controls at run time that aren’t driven by a "look pose" but rather driven by the joint, and thus takes the fleshy eye corrective morphs into account at full fidelity. I expect we will talk about this further once we’ve rolled out the system to every face, but it’s something quite novel and very powerful.

Consider that this for the base feature set reduces memory footprint by a factor of N, where N is the number of heads you have.This is the definition of a highly scalable system. One other huge win that we hope to garner from the system is customization that allows the player to adapt the control joints on the rigs. We then write out to the player persistence these offsets (in a clever manner) and we can then drive that fully customized head in the same way and without needing any bespoke animations.

The process for converting the heads is not an overly difficult one, but anytime we are making massive changes to 100+ head assets, we like to do this carefully. I expect we’ll demo this very soon.
 
Resident Evil 7 CEDEC Presentation- 659,254 primitives, 567.56 used memory.

vlcsnap-2016-09-11-00a9ure.png

vlcsnap-2016-09-11-00l2ud3.png
 
Last edited:
Well, it's a grimy first person horror game, and as far as I know it's also devoid of cut-scenes. I doubt we'll get plenty of good looks at any of the characters.
 
They poured everything into the model and left nothing for the environment which explains why all the trailers looked so sparse. Also the unrelenting use of CA and noise filters absolutely destroys any intricate details that may otherwise be shown clearly.
 
I saw Kingsglaive. Visual Works has outdid themselves, not only in rendering quality but also with script, pacing and action. Really enjoyable CGI film and a new benchmark for future realtime games. :)
 
"I think its not helpful for people to see floating heads ;) The biggest thing that will pull you out of it is the lack of body motion. We don't see people often PERFECTLY still and just moving areas on the face all isolated."

That is a very good point! All our face rigs have an ortographic camera parented to the head joint, so it keeps the entire head static and that allows for some interesting sights.

Even the individual Action Unit shapes (based on FACS) have a lot of movement; and once you start combining effects of multiple facial muscles which pull the skin in different directions, the amount of wobbliness for something as simple as talking can be quite disturbing at first. Basically everything under the brow bone is relatively thick soft tissue, including the tip of the nose; and also, skin deformations tend to have a large area as it prefers to slide instead of stretching.

Then there are the secondary effects that most CG face rigs don't really deal with - gravity and inertia. Now that's some really scary stuff :D
 
I watched the Playstation Meeting 2016 again on YouTube and looked carefully at all the characters in the game videos. Sadly, none of them impressed me. It seems like almost every game developer is keeping their characters stylized rather than going for photorealism. I was hoping there might be at least one game that tried to push the realism of characters, but there wasn't one. And in some games like Watch Dogs and Spiderman, the stylism of characters (for example pedestrians on the street) was very high. I realize in open world games that it would take a monster budget to try to make every pedestrian on the street look even semi-realistic, but there doesn't seem to be any effort going into making them look better. Now, when it comes to other aspects of gaming, such as draw distance, the number of vehicles on screen at once, the amount of "stuff" in the environment, the density of plants/bushes, there have been significant improvements.
 
I watched the Playstation Meeting 2016 again on YouTube and looked carefully at all the characters in the game videos. Sadly, none of them impressed me. It seems like almost every game developer is keeping their characters stylized rather than going for photorealism. I was hoping there might be at least one game that tried to push the realism of characters, but there wasn't one.
Why do you think that is?
 
In the latest 10 for the Chairman Episode they talked 30 minutes about their characters and tech (Star Citizen/Squadron 42).

Parts are still WIP. They admit that their hair is right now the weakest point and they want to improve the hair shaders to the best ones in realtime games. Eye shading and skin shading are also not done now.
They scanned 160 heads.

10 for the Chairman TLDR (Too Long;Didn't Read)
[1:23] Q: Will facial customisation be static and can only choose from a preset from heads or can we manipulate certain aspects using morphology.

A: You will be able to choose from a head that you like and from there you can change the skin colour, eye colour, hairstyle and some minor alterations to the physical shape.

[3:38] Q: How customisable will hairstyles be?

A: They will be fairly customisable, you can choose from an array of styles and change their colour, but you won't be able to alter their shape specifically, but you can choose to have a hairstyle on any head you’d want.

[5:10] Q: Can I reproduce my likeness in the character creator?

A: You should be able to get pretty close.

[5:51] Q: How long does it take to model someone’s face in the game?

A: A long iterative process that includes up to two hours for just the basic scanning of the face plus all the post processing work dealing with the data and multiple maps and shaders.

[11:27] Q: Backers faces in the game, is it possible?

A: They’ve considered the idea, talked about it but to be determined. They would want all the characters to have the same level of fidelity as any of the top tier characters and any home photogrammetry might give them information/textures that might not be up to standard of the quality they need.

[14:11] Q: How soon will the new heads be implemented and how editable will they be?

A: Not in 2.5, but it will be iteratively implemented along the way afterwards. We’re very close now.

[16:30] Q: What quality of faces will we see on our characters, NPCs, other players?

A: What you saw in the Morrow tour, forget that quality existed as they’ve moved on and have improved the fidelity. They have tier 0,1,2,3 characters and the character you saw in the Pupil to Planet was a tier 2 character. A tier 2,3 character will hold up to a tier 0 character because it’s about the small details and animations in the tier 0,1 characters that makes them so detailed. The lowest tier is 10,000 polygons and the highest is 40,000, but the higher quality tier’s allow the full range of the actors emotions such as Mark Hamill, Gary Oldman, Mark Strong to be displayed in the game which is why they’re a tier 0 and others are 1,2,3.

[21:07] Q: Will animations adapt to your character as they get older/recover from injuries?

A: They’re planning to do animations specific to injuries, want to have the impact of an injury reflect on the character’s animation. It’s a really cool idea to have older person locomotion and they have younger ones so it may not be that difficult to move you through those sets but they don’t have the inputs yet.

[22:31] Q: With character fidelity, how do you plan on keeping the resource cost down?

A: They’ve made huge strides so far. For example a character used to be 300 megabytes in memory, and now it’s down to 30 megabytes and the animation joints went from 1,000 to 183. There’s still more to be done, but they’re working on other ways to bring costs down further without losing fidelity.

[26:18] Q: Which character design approach is Star citizen using? Traditional where you choose from a number of presets or being able to edit vertex points within certain ranges or whole new approach?

A: Mix of both, start with a preset face and you choose face/hair colour/eye colour/etc and then mix and match from a combination of heads and you can move things around but not actually bringing in another mesh or other character information.

Source: http://imperialnews.network/2016/09/10-for-the-chairman-special-edition/

I' ve collected the clips from the video which were in my opinion the most important ones.
 
Last edited:
Very impressive, setting aside the creepiness factor. I want to see what they do with the hair, most models I've seen from this game have painted on textures rather than actual 3D planes and then using alphas, looks a bit disappointing when the actual models are so impressive.
 
Status
Not open for further replies.
Back
Top