Sensitivity as an Arbitrary Output Variable for Differentiable Rendering

arXiv cs.GR details a sensitivity AOV for differentiable rendering, framing derivatives as something artists and engineers can inspect like a rendered pass instead of a hidden byproduct of optimization. The paper by Linas Beresna and Eugene Fiume introduces a sensitivity buffer populated by a single reverse-mode pass, then reused for multiple views of the same data.

The practical angle is debugging and analysis. Developers can inspect image-space sensitivity at object and parameter granularity, project it onto a freely navigable scene from different viewpoints, and even carry per-texel fields through texture coordinates for spatially varying parameters. That makes it easier to understand which scene inputs actually drive a loss, especially when tuning inverse rendering or differentiable asset workflows.

The authors also separate the fixed camera that defines the objective from the free camera used to inspect the result, which is a useful mental model for anyone building tooling around differentiable pipelines. Rather than treating gradients as a one-off computation, the work positions derivative outputs as a render product alongside the primal image. The paper is listed as a SIGGRAPH Asia 2026 Technical Communications submission.