Summary and Team Members
Project Summary: We aim to reduce noise in Monte Carlo path tracing by combining Multiple Importance Sampling (MIS) with ML-guided adaptive sampling. MIS blends light and BSDF sampling strategies using heuristics, while the ML component uses saliency maps to guide adaptive ray allocation to perceptually important image regions. This will enable faster, cleaner rendering in regions that matter most to the human viewer.
Team Members: Eduardo Cortes, Yuhe Qin, Zhehao Yang, and Henry Michaelson
Problem Description
Path tracing produces photorealistic results but suffers from high variance (noise), especially in scenes with complex lighting and materials. Multiple Importance Sampling (MIS) mitigates this by combining BSDF and light sampling, each of which performs better under different conditions. Meanwhile, adaptive sampling dynamically concentrates effort where more samples are needed, often driven by variance or perceptual importance.
We propose combining these ideas with a lightweight ML saliency map to drive sample distribution. The intuition is to make semantically important regions (e.g., faces, shiny surfaces, object boundaries) converge faster by both allocating more rays and better balancing sampling strategies using MIS. We will try to ensure that this hybrid system can outperform either MIS or adaptive sampling alone.
Goals and Deliverables
Planned Deliverables
- Implement essential BSDF sampling functions to be able to represent complex objects
- Implement BSDF importance sampling and light sampling
- Combine them using MIS with power/balance heuristics
- Implement and incorporate ML models into adaptive sampling to probabilistically send rays to dynamic regions
- Compare performance against benchmark performance across the multiple strategies that we implemented for a variety of different final images
Hoped-for Deliverables
- Extend path tracing repository and base .dae files with more complicated scenes to render
- Use ML to adaptively shift sampling weight toward BSDF or light based on surface properties
- Develop visualizations of saliency and sample allocation maps
Schedule
- Week 1: Pivot idea from previous solution that was not valid as a final project. Create starter repository across team.
- Week 2: Implement BSDF functions, implement simple MIS using assets from Homework 3, create milestone deliverables.
- Week 3: Finalize MIS implementation, implement ML driven adaptive sampling, create benchmark of different solutions, acquire more complex .dae scenes to render (if possible), and ML integration into MIS (if possible).
- Week 4: Cleanup of materials for final presentation on 8/13.
Resources
References:
- PBRT Book: Importance Sampling and MIS
- MIS Blog (Chinese, via translation)
- CLIP by OpenAI (for saliency)
Tools & Frameworks:
- C++ for path tracer extensions
- PyTorch / OpenCV for saliency maps
- HTML/CSS for webpage
Code Repository:
Public Webpage: