This repository contains the code for the SIGGRAPH 2025 paper:

Reservoir Splatting for Temporal Path Resampling and Motion Blur
Jeffrey Liu (University of Illinois Urbana-Champaign), Daqi Lin (NVIDIA), Markus Kettunen (NVIDIA), Chris Wyman (NVIDIA), Ravi Ramamoorthi (NVIDIA and University of California San Diego)

The code is implemented as a standalone ReSTIR path tracer on top of the Falcor 8.0 framework [Kallweit et al. 2022]. Reservoir splatting is an extension of GRIS [Lin et al. 2022] and Area ReSTIR [Zhang et al. 2024], which forward-reprojects primary hits from the prior frame to the current frame. This preserves exact primary hits across frames, which generally improves the robustness of temporal resampling.

• All relevant code can be found in the Source/RenderPasses/ReservoirSplatting render pass.
• The script scripts/ReservoirSplatting.py can be used to set up the render graph.
• Falcor loads the pyscene format, which can import different scene formats.
  • Check out media/test_scenes/cornell_box.pyscene for a simple scene, but fancier scenes can be imported through Benedikt's rendering resources or pbrt-v4's scenes!
  • This implementation uses the reconnection shift after the primary hit, so delta materials are not supported. Refer to the GRIS paper [Lin et al. 2022] or the ReSTIR Course Notes [Wyman et al. 2023] for more details on different ReSTIR shift mappings.

Note that most of the reservoir splatting code has not been thoroughly optimized. Contributions are welcome!

Make sure to clone the repository with its required submodules:

git clone --recursive https://github.com/Jebbly/Reservoir-Splatting.git

For hardware / software prerequisites and instructions on how to build Falcor, check out Falcor's documentation.

After building and running Mogwai, run the script in scripts/ReservoirSplatting.py to set up the reservoir splatting render graph.

Depth of field and motion blur depend on the camera settings. When a scene is loaded, these can be enabled through Scene Settings -> Camera:

• A non-zero aperture radius will enable depth of field.
• A non-zero shutter speed will enable motion blur.
  • Since actual frame rate is variable, we use an artificial time between frames to compute motion blur. For example, if the shutter speed is set to 1/24 s and the artificial frame time is 1/48 s, motion blur is computed across half the frame time (even if the actual frame took less time to render).
  • You can change the simulated time between frames (defaults to 60 FPS) under ReservoirSplatting -> Rendering Options -> Simulated Frame Time.

The script defaults to splat-only temporal resampling, but different options can be enabled through the UI or by modifying the script. In the UI, under ReservoirSplatting -> ReSTIR Options -> Temporal Resampling, the following temporal resampling options are supported:

• GatherOnly is our (un-optimized) reimplementation of Area ReSTIR [Zhang et al. 2024].
  • Fast and Robust are the two options presented in Area ReSTIR, while Clamped uses integer motion vectors.
  • The robust reuse optimization is enabled by default, when the Robust option is selected.
  • If the camera has a non-zero aperture radius, both the lens vertex copy and the primary hit reconnection are used.
• ScatterOnly is our single-splat implementation.
• ScatterBackup is our splat + backup implementation.
  • Our results use Clamped as the gather method and Balance as the MIS weighting scheme.
• MultiSplatting is our multi-splat implementation.
  • This option is only available when the camera has a non-zero shutter speed.
  • If the partition count is set too high, your GPU may run out of memory.

We provide a set of utility in the UI for debugging and/or capturing output.

• Using Scene Settings -> Freeze, you can pause at your current frame and preserve the temporal history from the previous frame.
  • This is helpful for debugging / capturing under motion.
  • A reference image under motion can be captured by disabling all ReSTIR options and enabling the AccumulationPass. This is used for motion blurred reference images.
• Using Scene Settings -> User Interaction:
  • You can record the camera motion to a text file, and then replay the same motion from the output text file.
  • By default with Freeze Frame ID set to -1, the recorded motion will loop.
  • Setting Freeze Frame ID to a value greater than -1 will freeze the frame at the specified frame ID.
    • Unfreeze to continue playing the motion!
    • If Replay for Accumulation is enabled, the frame will no longer freeze but will instead contribute to the AccumulationPass. Enable View Accumulated Image and AccumulatePass -> Enabled to view the accumulation under motion, at the specified frame ID. This is useful for verifying unbiasedness under motion, but note that this is a very slow process since it effectively replays the entire motion sequence for one accumulated sample.
• Use the FrameDumper pass at the end to output every frame, which can be used to capture a sequence of frames and/or generate a video.
  • Warning: if you want to generate a video, disable Save PNG Files for better performance, and use FFmpeg.

A video can be generated as follows:

ffmpeg.exe -framerate 30 -i "C:/frameDump/frame.%05d.pam" -c:v h264_nvenc -preset slow -b:v 4M -pix_fmt yuv420p output.mp4

Note that higher bitrates are usually necessary to capture the noise characteristics properly.