Research map / Image generation / Fast diffusion sampling
RL-based Stateful Neural Adaptive Sampling and Denoising for Real-Time Path Tracing
Antoine Scardigli, Lukas Cavigelli, Lorenz K Muller
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 |
|---|---|---|---|
| train sampling network with RL instead of gradient approximation (ablation) | ours · Averaged over 4 scenes, 4 spp | PSNR (dB): 28.4 -> 28.7 | Improved |
| adaptive sampling instead of uniform sampling (ablation) | ours · Averaged over 4 scenes, 4 spp | PSNR (dB): 27.9 -> 28.7 | Improved |
| add latent space encoder with warped latent temporal feedback (ablation) | ours · Averaged over 4 scenes, 4 spp | PSNR (dB): 28.0 -> 28.7 | Improved |
| add temporal feedback and latent encoder (vs none) (ablation) | ours · Averaged over 4 scenes, 4 spp | PSNR (dB): 27.4 -> 28.7 | Improved |
| keep individual sampled pixel values instead of averaging (ablation) | ours · Averaged over 4 scenes, 4 spp | PSNR (dB): 28.2 -> 28.7 | Improved |
| scale down all three networks (ablation) | ours · Averaged over 4 scenes, 4 spp | PSNR (dB): 28.7 -> 27.3 | Hurt |
| use individual sampled values and latent encoder with gradient-approx sampling vs NTAS | ours-A1 vs NTAS · Averaged over 4 scenes, 4 spp | PSNR (dB): 27.7 -> 28.4 | Improved |
| learned state encoder vs exponential-average heuristic | ours-A2 vs IMCD · Averaged over 4 scenes, 4 spp | PSNR (dB): 27.9 -> 27.9 | No clear effect |
| learned sampling and latent update vs probabilistic heuristics | ours vs ReSTIR+OIDN · Averaged over 4 scenes, 4 spp | PSNR (dB): 28.0 -> 28.7 | Improved |
| use 2 spp budget vs strongest baseline at 4 spp | ours · Averaged over 4 scenes, 4 spp | PSNR and latency: 0.2 dB higher PSNR, 1.6x latency reduction | Improved |
| equal 4 spp budget vs strongest baseline | ours · Averaged over 4 scenes, 4 spp | PSNR (dB): +0.6 dB for 3 ms (5.5%) latency increase | Improved |
| shrink the denoiser network (ablation) | ours · EmeraldPlace, 100 ms budget | PSNR (dB) | Hurt |
Closest papers by meaning
- High-Quality Real-Time Rendering Using Subpixel Sampling Reconstruction (AAAI 2024 · similarity 0.89)
- Real-Time Neural Denoising with Render-Aware Knowledge Distillation (AAAI 2025 · similarity 0.88)
- Denoising, Fast and Slow: Difficulty-Aware Adaptive Sampling for Image Generation (CVPR 2026 · similarity 0.88)
- Learning to Sample From Diffusion Models via Inverse Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.86)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.86)
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.
