← All resources

Forget Superresolution, Sample Adaptively (when Path Tracing)

Auto-discovered

Paper Advanced

Martin Bálint, Corentin Salaün, Hans-Peter Seidel, Karol Myszkowski

Visit resource ↗

Description

Forget Superresolution, Sample Adaptively (when Path Tracing) presents an MPI for Informatics paper by Martin Bálint, Corentin Salaün, Hans-Peter Seidel, and Karol Myszkowski on end-to-end adaptive sampling and denoising for path tracing below one sample per pixel.

The method uses stochastic sample placement for trainable gradients at sparse budgets, tonemapping-aware perceptual training with filmic operators and MILO loss, and a gather-based pyramidal denoiser with learnable albedo demodulation. Results improve shadow boundaries and specular detail versus uniform sparse sampling and baselines such as DLSS Ray Reconstruction and NPPD.

It is for advanced researchers and engine developers working on neural denoisers and sample allocation for real-time path tracers.

Access is via the free arXiv abstract page and downloadable PDF preprint.

Adaptive sampling remains competitive when noise and perceptual importance vary across the frame. Link last verified accessible on 2026-09-27. Link last verified accessible on 2026-09-28.