Tobias Zirr - PowerPoint Presentation

Slide1

Tobias Zirr (1),2 and Anton Kaplanyan 11 NVIDIA Research2 Karlsruhe Institute of Technology

Real-time Rendering of Procedural Multiscale MaterialsSlide2

Previous Work“Sparkly but not too Sparkly! A Stable and Robust Procedural Sparkle Effect” [Studio Gobo, SIG AiRTR 15]Procedural noiseDirect placement of individual sparkle pointsPurely phenomenological2Slide3

Microfacet TheoryStochastic model for rough microscopic surface structure3

Normal Distribution Function (NDF):

 Slide4

Microfacet TheoryStochastic model for rough microscopic surface structure4

 Slide5

Microfacet TheoryStochastic model for rough microscopic surface structure5

 Slide6

Stochastic Models and CorrelationStochastic model assumes independent distributions6

More this:

Less this:Slide7

Stochastic Models and CorrelationStochastic model assumes independent distributions7

More this:

Less this:

Glint:

Local

aggregation of similar microsurface orientationsSlide8

Previous Work“Rendering Glints on High-Resolution Normal-Mapped Specular Surfaces” [Yan et al.14]Enumeration of reflecting ‘facets’ (texels) in footprintOrientation bounding hierarchy on normal mapTree pruning8Slide9

Previous Work“Discrete Stochastic Microfacet Models” [Jakob et al.14]Distribute arbitrary number of microfacets / ‘particles’Pseudo-random (deterministic), no need to store4D hierarchical stochastic distribution process (p, n)9

9Slide10

A Biscale Microfacet ModelIn Theory10Slide11

Multiscale Microfacets: MicrodetailsGeneral idea: NDF varies with locationModel microsurface by distribution of microdetails:“Randomly instantiate microdetail patches”11

Microdetails:

Randomized Instantiations of Microdetail NDFs/

“

MicroBRDFs

”Slide12

Relationship NDF, MDDF, MNDFEach microdetail represents a local microsurface with Microdetail NDF (MNDF):

Microdetails

are randomly oriented according toMicrodetail Distribution Function (MDDF):

The global microsurface consists of all microdetails with

NDF

:

Can decompose

any

NDF

into local and global DFs

to introduce correlated

clusters/glints!

12Slide13

Comparison to Previous Work“Almost Specular” MicrofacetsCount the number of reflecting microfacets per area[Jakob et al.]: Discrete Stochastic Microfacet Models ,[Yan et al.]: Rendering Glints on High-Resolution Normal-Mapped Specular Surfaces

Ours: Glossy MicrodetailsContinuum from max to min reflectivity per microdetail

13Slide14

Discretization of Microdetails“Almost Specular” MicrofacetsCount the number of reflecting microfacets per area[Jakob et al.]: Discrete Stochastic Microfacet Models ,[Yan et al.]: Rendering Glints on High-Resolution Normal-Mapped Specular Surfaces

Ours: Glossy MicrodetailsContinuum from max to min reflectivity per microdetail

Need to turn into discrete stochastic processTurn reflectivity of

into

probability

of

contributing with maximum (MNDF) intensity

14Slide15

Discrete Biscale Model EvaluationProbability

of

contributing with max intensity

Total probability mass

Expected number of contributing microdetails:

15

 Slide16

Discrete Biscale Model EvaluationProbability

of

contributing with max intensity

Total probability mass

Expected number of contributing

microdetails

:

Multiply contributing fraction

by

:

Locally has dynamic range MNDF

(zoomed in)

Expected value equals global NDF

(zoomed out)

16

 Slide17

Controlling Biscale NDFsPowerful artistic control:Local roughness / controls detail appearance

Global roughness /

controls distant appearance

(Microdetail distribution

implicitly defined by both)

17Slide18

A simplified proceduralreal-time ImplementationIn Practice18Slide19

Simplified Stochastic ProcessHierarchical multinomial process too expensiveSimplify to one binomial random variableTotal number of microdetails for a given areaProbability of microdetail contributing w/ max intensity19

 Slide20

Coherent Stochastic ProcessIdeally one binomial draw per pixel, but footprints varyPer-pixel area in screen space unstable!Resort to stable texture-space power-of-two grids and proven methods of anisotropic texture filtering:One binomial draw per grid cellTrilinear interpolation20Slide21

View Dependency (Glistening!)21

Search space 4D:

Also need subdivision of microdetail orientations

Paraboloid half vector grid

Seed binomial using 4D index

Perturb half vector partitioning

using texture grid index to avoid

simultaneous change of sparklesSlide22

Multiscale Coherent Noise / SeedsSeed binomials with cell indicesProblem: Blending noise leads to smearing (averaging)Solution: Multi-level coherent seeds22(Generic multiscale noise also useful elsewhere!)Slide23

Multiscale Coherent Noise / SeedsSeed binomials with cell indicesProblem: Blending noise leads to smearing (averaging)Solution: Multi-level coherent seeds23

cellIdx

>>

findLSB

(

cellIdx

)

(Generic

multiscale

noise also useful elsewhere!)Slide24

Anisotropic MicrodetailsAnisotropic roughness equivalent toanisotropic scaling of BSDF by ratio

“

Understanding

the masking-shadowing function in

microfacet-based BRDFs” [Heitz14]

Scale texture-space grid accordingly 

24Slide25

Microdetail Scale VariationVary microdetail density per texture grid cellSpreads grainy appearance across larger range of scales25Slide26

PerformanceGeForce GTX 980, 1080pMaximum anisotropy: 16xALU variance: 8-64 cells to shade412 static instructions, 204 within a loop for one cellNo texture fetches26

Scene

Polys

Isotropic footprint, ms

Grazing angle, ms

Full-screen pass

20.92.9Snow

32k

2.5

4.0

Dress

100k

1.4

4.4

Car (grooves)

570k

2.5

3.9

Crytek

Sponza

262k

3.0

5.9Slide27

Example Code / ShaderToyExample available online: https://www.shadertoy.com/view/ldVGRhSlide28

Variety of MaterialsSlide29

Thank you!Questions?ContactTobias Zirrtobias.zirr@alphanew.net Twitter: @alphanewAnton Kaplanyankaplanyan@gmail.com Twitter: @kaplanyan29