RL-based Stateful Neural Adaptive Sampling and Denoising for Real-Time Path Tracing
Authors: Antoine Scardigli, Lukas Cavigelli, Lorenz K. Müller
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 3.1 Experimental Setup, 3.2 Quantitative Results, 3.3 Qualitative Results, Table 1: Description, visual quality evaluation at 4.0 spp budget, and inference time in ms. |
| Researcher Affiliation | Industry | Lukas Cavigelli, Lorenz K. Müller Computing Systems Lab, Huawei Zurich Research Center, Switzerland and {lukas.cavigelli, lorenz.mueller}@huawei.com |
| Pseudocode | No | The paper describes the algorithms for its three models (Sampling Importance Network, Latent Encoder Network, Denoiser) in Section 2.3, but it does not present them in a structured pseudocode or algorithm block format. |
| Open Source Code | Yes | We release our dataset scenes and code implementation1 to facilitate the usage and comparison of our method. 1Source code available at https://github.com/AJSVB/RL_PATH_TRACING. |
| Open Datasets | Yes | We use three 3D scenes Emerald Square [40], Sun Temple [41], and Zero-Day [42] released as part of the Open Research Content Archive (ORCA) under CC BY 4.0 license. |
| Dataset Splits | No | The paper mentions 'cross-validated results for each scene: For every test scene, we train on all the other scenes', which describes a cross-validation strategy but does not specify explicit training/validation/test dataset splits (percentages or counts) for a single dataset. |
| Hardware Specification | Yes | We perform our measurements on a Nvidia Ge Force RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper states 'We use Adam optimizer [38], Pytorch and Ray-RLlib [39]' but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | All networks are trained in a closed loop on 100 epochs with a batch size of 4. ... the learning rate has a maximum value of 0.1 and a minimum value of 10 8, with a warmup phase of 15%, and then an exponential decay phase. ... We transform the images by using vertical and/or horizontal flips, by randomly cropping and rescaling the images. |