I must admit it's speculation on my part but it looks exactly like all the facegen + upscaling tech that Nvidia has published papers about for the last 10 years. Similar outputs are endemic to all the most widely used AI face models, explaining why DLSS5 looks like the fake ads I see on YouTube

the model is hallucinating new light sources on the face because it’s trained on upscaled portrait photography. the RE example is running path tracing. it is not using an insane hero lighting rig like what you’d need if you _wanted_ to produce something similar to the dlss version.

REAC 2025 talk details have been posted on the website! Register today - it's free! enginearchitecture.org/2025.htm

If you can share this message, we'd appreciate it. Thanks, and see you soon!

Measuring acceleration structures Hardware accelerated raytracing, as supported by DirectX 12 and Vulkan, relies on an abstract data structure that stores scene geometry, known as “acceleration structure” and often referred to as “BVH...

New blog post! "Measuring acceleration structures", in which we will compare BVH costs on various GPU architectures and drivers and attempt to understand the details enough on AMD hardware to make sense of the numbers!

Reposts appreciated :)

zeux.io/2025/03/31/m...