Research map / Optimization / Optimizers
Stochastic Gradient Estimation for Higher-Order Differentiable Rendering
Zican Wang, Michael Fischer, Tobias Ritschel
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use aggregate importance sampling for the Hessian-vector product (ablation) | OurHVP · Texture (1024 parameters) | relative convergence time: 4.00 -> 1.00 | Improved |
| use aggregate importance sampling for the Hessian-vector product (ablation) | OurHVP · CNN5 | relative time to .9 image error: 3.64 -> 1.00 | Improved |
| use aggregate importance sampling for the Hessian-vector product (ablation) | OurHVP · Mug (single parameter) | relative time to .99 image error: 1.00 vs 1.14 (OurHVPA not faster) | Hurt |
| use aggregate importance sampling (ablation) | OurHVP · Box2 (2 dimensions) | relative time to .99 image error: 0.64 vs 1.00 | Hurt |
| use HVP instead of full Hessian (ablation) | OurH · Texture | relative time to .9 image error: 40.12 -> 4.01 | Improved |
| use HVP instead of full Hessian (ablation) | OurH · Box10 | relative time to .99 image error: 5.33 -> 1.44 | Improved |
| use second-order Hessian estimate instead of first-order smoothed gradient | OurG · Mug | relative time to .99 image error: 3.91 -> 1.00 | Improved |
| use our importance-sampled smoothed gradient instead of FR22 | FR22 · all tasks | convergence time | No clear effect |
| replace gradient descent by higher-order optimization with OurHVPA | gradient descent baselines · all tasks | wall-clock convergence speedup: 2.71x average | Improved |
| use the Mitsuba analytic gradient | Mitsuba · Mug, Shadow, Quad, Box, Texture, CNN | convergence: does not converge on plateaus | Hurt |
| increase rendering samples 0.25x to 3x | OurHVPA · Mug | optimization quality and convergence | No clear effect |
| higher-order method on higher-dimensional problem | OurHVPA · Texture and CNN | convergence: performance delta lessens with dimensionality | Mixed |
Closest papers by meaning
- Locally Orderless Images for Optimization in Differentiable Rendering (CVPR 2025 · similarity 0.87)
- Diffusing Differentiable Representations (NeurIPS 2024 · similarity 0.86)
- High-Quality Real-Time Rendering Using Subpixel Sampling Reconstruction (AAAI 2024 · similarity 0.86)
- Diffusion Renderer: Neural Inverse and Forward Rendering with Video Diffusion Models (CVPR 2025 · similarity 0.85)
- Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs (CVPR 2025 · similarity 0.85)
- Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
