Tzu-Chieh Chang

I love rendering (:
Opinions are my own.

[Tzu-Chieh Chang @TzuChieh_Chang]

@anki3d Interesting read. Given the result, I think even ray tracing pipeline can make use of "potato RT" after a certain bounces.

[Tzu-Chieh Chang @TzuChieh_Chang]

@LunarGInc This seems wrong. Someday we may see RenderDoc in Vulkan SDK 🤔

[Tzu-Chieh Chang @TzuChieh_Chang]

Most BSDFs in the room are blended between diffuse and glossy, while photon mapping does not play well with non-diffuse BSDF. To overcome this, I added BSDF lobe splitting, so that vertex merging can still be performed if the BSDF contains at least 1 diffuse lobe. It works well!

[Tzu-Chieh Chang @TzuChieh_Chang]

@EskilSteenberg @quelsolaar Wow, never noticed Hyperion is also a moon. That renderer is a beast.

[Tzu-Chieh Chang @TzuChieh_Chang]

There are too many sampling strategies in my renderer that I am no longer confident in adding a new strategy without breaking old ones. A chi^2 test for goodness of fit is useful in this case and has helped me finding a new bug in dielectric sampling code, even after 10 years...

[Tzu-Chieh Chang @TzuChieh_Chang]

@elonmusk First product should be Macrohard Onfire, AI assisted variant of Microsoft Office.

[Tzu-Chieh Chang @TzuChieh_Chang]

@vfxpro88 Unfortunately this is an offline render. I do believe this can be made real-time though. It's cost is basically an extra microfacet BSDF evaluation.

[Tzu-Chieh Chang @TzuChieh_Chang]

Implemented BSDF evaluation and PDF query for microfacet-based normal mapping to support more light transport algorithms. For dielectrics, I need to make some adjustments to the equations in the original paper. Now I need to find a beautiful scene and test it.😉

[Tzu-Chieh Chang @TzuChieh_Chang]

@tigratigratigra Sure, here it is
https://t.co/msKXcaX7Nq

[Tzu-Chieh Chang @TzuChieh_Chang]

@turtlespook Regular normal mapping simply replace the normal with a tabulated one, so it's not physically based. Microfacet-based normal mapping works like a lenticular card--the surface appears bumpy because there are tiny facets facing a different direction from the geometric normal.

[Tzu-Chieh Chang @TzuChieh_Chang]

@tigratigratigra The normal map is from their supplemental material https://t.co/MzBbuAU2Gz
The rest is in my custom scene format. Walls: 10x10, gold (GGX, alpha=0.16), white (Lambertian diffuse, albedo=0.7); glass ball: r=2.5 (IoR=1.5); area light: 4x4, tilted 30 degrees (500 Watts).

[Tzu-Chieh Chang @TzuChieh_Chang]

@tigratigratigra Hi, it's "Microfacet-based normal mapping for robust Monte Carlo path tracing"
https://t.co/xzg7NpYX6L

[Tzu-Chieh Chang @TzuChieh_Chang]

First sensible output with microfacet-based normal mapping in Photon-v2 renderer. If the geometry comes with flat normal/tangent, it can preserve symmetricity for bidirectional light transport out of the box.

[Tzu-Chieh Chang @TzuChieh_Chang]

Priority based nested dielectrics can also be used for disambiguating overlapping volumes. In the following image, the blue sphere has the highest priority, followed by green and red spheres. Medium properties are obtained from the volume with highest priority.

[Tzu-Chieh Chang @TzuChieh_Chang]

A high quality one. Start to see some geometric imperfection.

[Tzu-Chieh Chang @TzuChieh_Chang]

Tracking medium interfaces was not as easy as I originally thought. Anyway, this is the Utah teapot with simple dielectric material, attenuated by Beer's law. Next step will be trying to put some liquid in there 🥹

[Tzu-Chieh Chang @TzuChieh_Chang]

@clabaultt Yes, and the importance sampling algorithm they proposed is also quite effective.

[Tzu-Chieh Chang @TzuChieh_Chang]

Several years ago I implemented Laurent Belcour's layered BSDF, and its smooth appearance always makes me happy. Sadly, I still cannot reproduce the volumetric part.

[Tzu-Chieh Chang @TzuChieh_Chang]

@max_tarpini @clabaultt This is exactly what I was trying to solve. Concave primitives still cannot be resolved robustly by my method.

[Tzu-Chieh Chang @TzuChieh_Chang]

Testing my bisection-based self-intersection avoidance algorithm on an interior scene (40 spp). Overall, we see ~12% increase in execution time. On average there are 3 re-intersections per hit event (2 iterations). I think I'll take that... 🤔