Recent searches
Search options
Really excited to see how the DLSS API grows to provide this ray reconstruction layer (secondary motion vectors? mask? separate history submit?) and the practical upshot of it. Currently is clearly a weakness of all upscaling reconstruction techniques that becomes more important as more specular is provided by high quality RT going forward. https://videocardz.com/newz/nvidia-announces-dlss-3-5-with-ray-reconstruction-launches-this-fall
More details "incorporates additional engine data" (materials & geometry so looks like it probably wants to consume your fat g-buffer) & "high frequency sampled rays" showing independent internal-res history buffer for noisy RT samples. https://youtu.be/sGKCrcNsVzo
Seems like DLSS 3.5 RR is your GI solution (check out the significant different to CP77 example once you're using AI RR for that specular/bounce lighting rather than denoising such effect via current RT Overdrive mode).
Text version of video content: https://www.nvidia.com/en-us/geforce/news/nvidia-dlss-3-5-ray-reconstruction/
The more I've looked at this outline of the RR technique, the more it looks like a major departure from DLSS & into something a lot closer to the RTXDI+RTXGI complete lighting solution that nVidia are offering for PT in CP77 (etc).
Handing low res aliased fat g-buffer scene info & sparse ray results from which the DL infers the full scene lighting results (think the best denoiser you've ever seen, ingesting your g-buffer to aid reconstruction) while also doing full upscaling, frame-gen, & AA.