Great project, with ReSTIR it could be perfect

[qlee01]

Hi,

this looks fantastic and is superior to most Path Tracing implementations I have seen, especially because it only uses screen space / reflection probes and does not need to build a BVH. Are there any future plans for this? I'd guess adding ReSTIR would be prefect for this to enhance performance even more, allowing to run with 1 sample per pixel only. I tried to integrate, but failed so far.

[jiaozi158]

Hey, thanks for trying out this repo!

Are there any future plans for this?

ReSTIR is a powerful sampling algorithm, but since the lights in SSPT (initial rays) are rendered via rasterization, it may not be suitable for this repo AFAIK.

However, I know there are still some sampling improvements that could be considered:

1. Environment Importance Sampling

For any point on a surface, we can get its lighting conditions in different directions from reflection probe. (By sampling the probe along a given direction)

Darker areas require fewer samples. For example, if the reflection probe shows pure black (indicating no lighting in the scene), rendering just 1 sample produces the same result as 1,000 samples.

By using an environment importance map to bias the random sampling directions, we can reduce the samples needed for the image to converge.

Check this for details: Environment map importance sampling - Karim Naaji

2. Reuse nearby samples if valid (similar to ReSTIR)

I'm not familiar with its details yet, maybe I can look into it when I have time.

Hoping these improvements could be easy to implement in Unity.

... and does not need to build a BVH.

I don't have plans to add BVH tracing yet, but this is possible since Unity has already done the most challenging part for us. (Unified RT package)

One problem is that BVH only stores the UVs of hit points. So for compute-based ray tracing, we need to collect & update the objects' material data into texture atlases or arrays, which makes adding it more complicated.

[qlee01]

Thanks, this sounds good! I understood that the original ReSTIR would work with a sole screen space solution by doing spatial-temporal reuse of the reservoirs for each initial ray, while ReSTIR GI (which is a subsequent work) requires actual path tracing in geometry, but maybe I am wrong with this.

[qlee01]

So I did a bit more research, and actually it seems vice versa, ReSTIR GI should be a good fit. Will try again to implement, I did a first working port of this repo To HDRP, will try it there, as actually the API for custom passes seems way easier than Render Graph stuff for URP, and stuff like motion vectors are better supported. Also DLSS helps a lot in cleaning up the output.

[jiaozi158]

Hi, I've been busy with some modeling stuff lately, but excited to hear that adding ReSTIR is a viable direction.

I did a first working port of this repo To HDRP, will try it there, as actually the API for custom passes seems way easier than Render Graph stuff for URP, and stuff like motion vectors are better supported.

Totally agree, issues like motion vectors are at the engine level, and for some reason (BiRP compatibility?) the HDRP team could only fix them in their package.

Hopefully these will be fixed in Unity 7!