Chris Wyman

Our upcoming HPG 2025 paper ”Collaborative Texture Filtering” is up! Research done with @_pandersson, Matt Pharr, and @BartWronsk. research.nvidia.com/labs/rtr/publi…

@BenSimsTech AFAIK, nobody is using it today. If they are, they're keeping quiet. I'd be willing to have an exploratory chat if anyone was interested, though I suspect it's the wrong choice in most situations. It suffered from code complexity in a way RT definitely does not.

Did frustum-traced shadows ever continue to be used/improved by any studios/engines since it was introduced in 2016 or is it dead tech because of raytracing? Seems like it could still be valuable since it can utilize raster hardware while RT happens async developer.download.nvidia.com/gameworks/even…

@SebAaltonen You should ask them on their forums / discord. They're usually pretty responsive. It might be nobody has asked them this question yet, so it's not documented.

Could somebody educate me on how the "half" type works in Slang? ChatGPT hallucinates way too much, and I can't find the info on their website. Questions: 1. Can I select fp16/fp32/relaxed output? 2. How half works in data interfaces (UBO/SSBO) with all above choices?

I3D 2025 will be held at New Jersey Institute of Technology, from the 7th to 9th of May! A couple of the dates of the paper submission process have been shifted. Find more details in the website: i3dsymposium.org/2025/cfp.html #I3D2025

The open source Slang Shading Language has transitioned from hosting at NVIDIA to Khronos under open governance! All now have an equal say in its evolution. Vulkan and many other APIs now have a portable, modern, fast-evolving Shading Language! khr.io/17f

It's getting to that point where our research team is thinking about next year's interns. If interested, feel free to e-mail me, or (even better) find someone on our team whose research aligns with yours and e-mail them. research.nvidia.com/labs/rtr/

We're looking for passionate innovators to join us in developing next-gen DLSS and revolutionizing graphics with AI once again! nvidia.wd5.myworkdayjobs.com/en-US/NVIDIAEx…

🌟Exciting News!🌟 The NVIDIA graduate fellowship is now accepting applications for this year's cycle! 🚀 The deadline to apply is September 13, 2025. For more details, check out this blog post: blogs.nvidia.com/blog/applicati… #NVIDIA #GraduateFellowship #Opportunity #Deadline #ApplyNow

I'm going to HPG and SIGGRAPH in Denver. E-mail or DM me if you want to meet to: catch up, talk, brainstorm, introduce yourself, etc.

Crazy request! I just ran into this position at Sony PlayStation : boards.greenhouse.io/sonyinteractiv… Among other things it's about raytracing, reservoirs, denoising, all the good stuff 😍. Would it be too much to ask for a referral/pointer to help me out? RTs appreciated!

@LukeMamacos Nice! Congrats -- it looks great! (And I have been accused of black magic before on ReSTIR, so welcome to the secret club!) Though it's not just my work; by this point it's mostly all my collaborators doing the hard work!

I fixed the TAA bug causing the black borders, so now it's just a bit noisy but a bit of hole filling or noise filtering should clear that up :> I just love this algorithm so much 🥰 I feel like I'm a wizard doing black magic. Thank you for your gift to the world @_cwyman_!

@LukeMamacos Everything else in the paper is reframing the math so we can postpone reconnections between the paths correctly. Basically ReSTIR GI uses one of the simplest shift maps around. ReSTIR PT uses one that's somewhat better designed.

@LukeMamacos ReSTIR GI connects the new pixel to the reused path immediately after the primary hit. ReSTIR PT realizes this may behave poorly since these surfaces may be highly specular and simply postpones reconnecting the paths for a bounce (or a few) until the surface is diffuse enough.

I realized that you can easily add the weighted lighting contribution of all spatial samples to the final output color while generating the reservoir that you feed back to the next frame. Before/After with/without TAA: #ReSTIR #Vulkan #RayTracing

@LukeMamacos Reflection & refraction are fairly easily handled by ReSTIR PT (but not ReSTIR GI). On glass, you probably want to choose between them based on Fresnel terms.

@_cwyman_ Ah, fair enough, PDFs for glass reflection/refraction are my current headache.

@LukeMamacos That's one of the ideas in our HPG paper: decoupled reuse and shading. (You reuse one sample and shade all.) It can be a big win, but there are cases where it doesn't help. There may also be cases where it makes things worse (though these are probably few).

@_cwyman_ I didn't see this mentioned in any of the papers, maybe I missed it.

@BattleAxeVR @NateMorrical This is a constant problem with people assuming magic denoisers. People forget that denoisers apply only to certain noise inputs. If you apply such a denoiser off-label, you run the risk of people saying "your sampling algorithm sucks" when that's likely not the problem.

@BattleAxeVR @NateMorrical Problem: Most off-the-shelf production real-time denoisers assume non-noisy G-buffers as guide buffers. Depth-of-field does not provide obviously usable inputs for these.

Area ReStir is absolutely gorgeous! What I wonder about is when this will be followed up with a bokeh or DoF-aware denoiser. I would kill to have this run at high framerates in VR. Maybe Blackwell? Or perhaps simply with DLSS + frame gen could possibly do it. Or multi-gpu.

@BattleAxeVR Sounds like a bug in Song's code. There's often strange limits in research code. Some of our Falcor path tracers have limit images to 4096 pixels wide. Here's an image from our internal implementation: 4072 x 1421 pix

Aw man, ok this is something I have to fire up my WOLED monitor to see. If you stretch the window beyond 1080p the app crashes! It complains it can't handle higher, even on my 4090! ouch. 1080p 30-40 fps range.

@svadrb @SongZhang__ But, yes, it's unclear if temporal reuse is a good thing offline. It's also unclear if (instead) you should be reusing samples from both past *and* future frames.

@svadrb @SongZhang__ Many things. Importantly: available time per frame. You might be able to spend 10sec or 10min per frame building a data structure that accelerates your rendering. You can't do that in real-time. This changes the usable set of algorithms, which might impact which is best.

@SongZhang__ 's work is a simple, but significant, improvement over prior ReSTIR algorithms. Definitely take a look at the paper! One sample per pixel antialiasing and depth of field, which also opens up quality improvements in a number of challenging situations.