Gracien

[Gracien @gracien_app]

Experimental UD Path-Tracer written in Swift, utilising #AppleMetal and #SwiftUI exclusively. M1 Pro

Microfacet BSDF showcase with minimal menu. Engine is controlled through a custom scene format.

Scene: Official Swatch from
@pixarsrenderman

https://t.co/9hw7tBd4sn

[Gracien @gracien_app]

@yiningkarlli It looks very native. Also love the attention to detail regarding metadata. I recall your previous tweets about "marrying" SwiftUI and CMake; avoiding pure-Xcode workflow; curious -- how'd that go?

[Daniel Lepik @lepikdaniel]

form explorations.26.1

_

[Gracien @gracien_app]

@yiningkarlli @Gok Sounds doable? https://t.co/h6Q2CcufqQ

[Gracien @gracien_app]

@novikoff Beautiful colour!

[Yining Karl Li @yiningkarlli]

The @pixarsrenderman team has a paper out at HPG 2025 this week about the architecture of RenderMan XPU. There's a lot of interesting details in the paper- definitely a worthwhile read!

https://t.co/inufBcB2W0

[Gracien @gracien_app]

@Xavyiy Metal 4 looks like a solid update... on all fronts 😁 and MSL going C++17!

[Gracien @gracien_app]

@VrKomarov Interesting, might check earlier versions, thanks!

[Gracien @gracien_app]

@VrKomarov I understand, my usage of the denoiser is not games-oriented either 🙂.

Thanks! I must say these are very surprising results. It's either the Metal backend or... version? I've only really tested OIDN 2.3.2+

[Gracien @gracien_app]

@VrKomarov Right... it should've added -- running Metal backend.
In your case, guessing it's CUDA or HIP?

[Gracien @gracien_app]

@VrKomarov When you say it's pretty fast, how fast is it 😁?
I'm seeing ~54ms in total on your end, what's the resolution of the denoiser-input image?
In Blender, on "Fast" setting it seems to take around >100ms for 1080p image (with pre-filtering of auxiliaries turned off).

[Gracien @gracien_app]

@yiningkarlli "Smena Lomo T-43 40mm ƒ/4"
Man, have to put my Smena 8M to use again 🙂

Great photos, squeezed all juice from that 5s sensor on the last photo 😁

[Gracien @gracien_app]

@yiningkarlli https://t.co/BX3w60FS8h
Sounds perfect for recreating this song 😁

[Gracien @gracien_app]

@PascalLecocq @yiningkarlli Waiting for the updates then :) Thanks for explaining!

[Gracien @gracien_app]

@yiningkarlli @PascalLecocq Had a hard time interpreting what the multiplier does; the default value of 1.0 gives around 15% improvement in "bouncier" :) scenes on average, whereas testing 0.1-0.9 sometimes hit a regression (20% at worst). I'd think it's Apple's SER but NV SER works fine (per Spear's paper)

[Gracien @gracien_app]

@yiningkarlli @PascalLecocq The artifacts are well pronounced using the Metal GPU backend in Cycles. Can't seem to test this reliably, not seeing much of a difference in performance either.

[Gracien @gracien_app]

@yiningkarlli By aligning the stochastic path termination (via similar base seed), they were able to get performance on par with the wavefront approach... a 2x gain sounds pretty significant - have you tried experimenting with this yourself?

[Gracien @gracien_app]

@yiningkarlli And they kept it in in the original format 😁
That's handy, thanks!

[Gracien @gracien_app]

@Pjbomb2 @clabaultt Interesting! Sounds like a pretty custom approach then? The MIS issue should still hold valid if you mix samples from different sources... anyways :)

You are now using Pairwise MIS, any difficulties / surprises while implementing? How did the perf change vs the 1/M weights?

[Gracien @gracien_app]

@Pjbomb2 @clabaultt If you combine different distributions then using 1/M is leading to bias, yeah. It's really a special case of the regular MIS weight where the other PDF is 0.

What i meant is generating candidates using not only direct light sampling, but also BSDF sampling.