Bart Wronski 🇺🇦🇵🇸

Path tracing is the future of video games rendering, no doubt. Not convinced?

Watch this video, the difference is stunning. https://www.youtube.com/watch?v=I-ORt8313Og

Every light and every single emissive polygon casting shadows and multi-bounce GI?
No problem, no hacks needed.

A ton of research and work from my Nvidia colleagues and folks at CDPR went into it.

Cyberpunk 2077 Ray Tracing: Overdrive Technology Preview on RTX 4090

YouTube
Apr 10, 2023 at 7:02pmWeb
Show threadJosh SimmonsApr 10, 2023
@BartWronski Too bad the exposure is completely broken so everything looks like a throwback to 2008 levels of bloom. :'(
Show threadStefan IvanovskiApr 10, 2023
@BartWronski How much more research would there need to be after real-time pathtracing becomes feasible for most games? How close would we get to having computer graphics as a whole being close to solved (for practical purposes)? Even though these amazing innovations have taken place and I'm amazed by the pace of research, I feel a bit discouraged as a newcomer to the field to be honest. What are your thoughts?
Show threadNathan ReedApr 10, 2023

@lmpxl @BartWronski there's still tons of research on light transport going on - path tracing itself is far from a solved problem! ReSTIR for example is a very recent innovation that has dramatically shaken up the field.

Even in a hypothetical "path tracing everywhere" future, there will still be a need for research toward: light transport perf and robustness, large scale scenes, streaming and LOD, geometry and material representations, volumetric media, animation and physics, etc...

Show threadNathan ReedApr 10, 2023
@lmpxl @BartWronski And then there's the ML side of things - I'm sure we've only barely begun to see the ways that differentiable programming and ML models can be advantageous in graphics and animation.
Show threadBart Wronski 🇺🇦🇵🇸Apr 10, 2023
@reedbeta @lmpxl everything you mentioned plus things like denoising path tracing (latest flavor of DLSS does an exceptional job on those 1spp images, but it can be improved for sure) :)
then things like path traced (real) motion blur and DoF, combining this with nanite-like geometry, infinite LOD etc. etc.
Show threadKrzysztof NarkowiczApr 11, 2023
@lmpxl @BartWronski IMO we still are at least one or two console gens (5-15y?) before pure ray tracing (zero raster passes) becomes viable in most games. Depending on how you define Path Tracing that will take even longer. Especially If you follow the original Path Tracer definition (no spatiotemporal ray reuse etc.) - https://www.pbr-book.org/3ed-2018/Light_Transport_I_Surface_Reflection/Path_Tracing
Path Tracing

Show threadBart Wronski 🇺🇦🇵🇸Apr 11, 2023
@k_narkowicz @lmpxl For the timeline, I agree, even when it comes just to path traced lighting, 5-10y because of consoles, workflows, time to adapt...
And as a side note - I don't even think "pure" path tracing without "any" rasterization is a goal (on its own). If it works well, fast, and flexible enough for primary/camera visibility, then why change it. :)
Show threadKrzysztof NarkowiczApr 11, 2023

@BartWronski @lmpxl What's your definition of path-traced lighting? As linked in my previous tweet, in the original definition there's no concept of lighting only or spatiotemporal ray reuse, just brute force rays. Unlikely that even on next-next-gen we can do that.

Regarding primary rays, it's not that useful on it's own, but it means that now we solved the hardest issue - entire scene is in BVH without any shortcuts! Which means ability to render without bespoke solutions/passes per feature!

Show threadBart Wronski 🇺🇦🇵🇸Apr 11, 2023
@k_narkowicz @lmpxl my definition of path-traced lighting would be varied length paths without distinction between them and a flexible decision to continue one or not. Whether the first vertex is at the camera or the first hard surface is irrelevant, IMO.
Show threadKrzysztof NarkowiczApr 11, 2023
@BartWronski @lmpxl Fair. Regardless of naming, my point is that until we'll have to rely on spatiotemporal path reuse, caching, guiding and other means for exploiting specific feature constraints there will be always ability to innovate and research. And solving lighting by "just trace more rays and denoise the end result" seems like a quite distant future.