Accelerated cross-platform implementation of differentiable rendering of signed distance function

Accelerated cross-platform implementation of differentiable rendering of signed distance function

Budak A.S. (KIAM RAS, Moscow, Russia)
Garifullin A.R. (KIAM RAS, Moscow, Russia)
Galaktionov V.A. (KIAM RAS, Moscow, Russia)
Frolov V.A. (KIAM RAS, Moscow, Russia)
Voloboy A.G. (KIAM RAS, Moscow, Russia)

Abstract

Reconstructing scene parameters from images (inverse rendering) is a popular problem and an important area in computer graphics and vision. Differentiable rendering, based on gradient optimization methods, is currently being increasingly applied to this task. This paper presents improvements to the method for differentiable rendering of signed distance functions, proposed in 2024, as well as a cross-platform implementation that supports execution on various types of graphics accelerators. This ensures independence from specific hardware vendors and expands the applicability of the method to heterogeneous hardware configurations. Our paper proposes two key modifications. First, we replace the standard ray tracing method with Newton's method and an analytical method adapted to differentiable rendering problems. Furthermore, we split the calculation of derivatives with respect to texture and geometric scene parameters into two parts, corresponding to the interior and boundary integrals. This partitioning reduces the number of Monte Carlo samples required to estimate texture gradients and allows the computation to be distributed between two shaders. As a result, the developed implementation of differentiable rendering is three times faster compared to the baseline implementation while maintaining the same level of accuracy.

Keywords

computer graphics; differentiable rendering; ray tracing; cross-platform.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 3, part 3, 2026, pp. 27-38

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(3)-33

For citation

Budak A.S., Garifullin A.R., Galaktionov V.A., Frolov V.A., Voloboy A.G. Accelerated cross-platform implementation of differentiable rendering of signed distance function. Proceedings of the Institute for System Programming, vol. 38, issue 3, part 3, 2026, pp. 27-38 DOI: 10.15514/ISPRAS-2026-38(3)-33.

Full text of the paper in pdf (in Russian) Back to the contents of the volume