ReFrame: Layer Caching for Accelerated Inference in Real-Time Rendering
Authors: Lufei Liu, Tor M. Aamodt
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that we achieve 1.4 speedup on average with negligible quality loss in three real-time rendering tasks. Code available: https://ubc-aamodt-group.github. io/reframe-layer-caching/ |
| Researcher Affiliation | Academia | 1Department of ECE, University of British Columbia, Vancouver, Canada. Correspondence to: Lufei Liu <EMAIL>, Tor M. Aamodt <EMAIL>. |
| Pseudocode | No | The paper describes the methods and architectural modifications in text and through diagrams (e.g., Figure 2, Figure 10). It does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code available: https://ubc-aamodt-group.github. io/reframe-layer-caching/ |
| Open Datasets | Yes | We generate our test set with the Unreal Engine build published by the authors and free scenes from the online Unreal Engine Marketplace (Fab), detailed in Appendix A.1. ... Authors of Implicit Depth (Watson et al., 2023) released their sample test scenes, which we use to evaluate the performance of Re Frame on their network. ... We compare the amount of motion in our test set to those commonly observed in videos of real-world game play from Gaming Video SET (Barman et al., 2018) and CGVSD (Zadtootaghaj et al., 2020). |
| Dataset Splits | No | The paper specifies the number of test frames and frame sequences for evaluation in Appendix A.1 (Table 6). However, it does not provide details on training, validation, or explicit splitting methodologies for these datasets, nor does it refer to standard dataset splits. |
| Hardware Specification | Yes | We demonstrate empirically on an NVIDIA RTX 2080 Ti GPU that using Re Frame reduces the inference time of three different real-time rendering networks... For latency, we compare relative speedup in inference latency measured on a NVIDIA RTX 2080 Ti GPU. |
| Software Dependencies | No | We modify each network in Py Torch to add our caching scheme, following implementation details in Appendix A.2. The paper mentions PyTorch but does not specify a version number. |
| Experiment Setup | Yes | We choose frame deltas with both high (Delta H) and low (Delta L) sensitivities as the default policy to refresh the cache for all experiments except the ablation study in Section 4.5, where we compare different cache policies. ... Table 7 shows the detailed settings of the cache refresh policies used in the ablation study. Configuration: Delta H (Parameters τ = 0.20), Delta L (Parameters τ = 0.25), Motion Vector (Parameters τ = 1), Non-Linear (Parameters c = 110, p = 1.4). |