Research map / Image generation / Fast diffusion sampling
Real-Time Neural Denoising with Render-Aware Knowledge Distillation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RAKD, a knowledge distillation framework for real-time Monte Carlo denoising that distills a kernel-prediction teacher into a lightweight student using GAN-based global loss, render-aware parameter initialization, and an auxiliary teacher-labeled unlabeled dataset.
Problem
Real-time Monte Carlo ray tracing at 1-4 spp needs a denoiser that balances quality and speed, but prior neural denoisers are either large and slow or lightweight with lower quality and over-blurring.
Approach
Applies knowledge distillation to MC denoising with three render-specific ingredients: pixel-level plus Wasserstein-GAN image-level distillation, initializing the student's U-Net by subsampling teacher kernel-predictor weights, and training on unlabeled frames whose noise-free targets are generated by the teacher.
Key results
- Student Ours (S) reaches 25.36 dB PSNR on Bistro at 4 spp vs 23.86 for ONND and 24.33 for NPPD, while running a full HD frame in 10 ms.
- Teacher Ours (T) reaches 26.54 dB on Bistro, above OIDN at 25.26 dB.
- Ablation: removing the GAN/global loss drops Classroom PSNR from 30.62 to 26.54 for the student without the discriminator.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the auxiliary unlabeled dataset to student training (ablation) | RAKD student · Bistro, Zero-Day, Classroom, Dining room | PSNR: 24.58 -> 25.36; 25.97 -> 26.95; 26.54 -> 30.62; 32.43 -> 35.02 | Helped |
| add GAN regularization and global loss (ablation) | RAKD student · Bistro, Zero-Day, Classroom, Dining room | PSNR: 24.93 -> 25.36; 26.69 -> 26.95; 29.64 -> 30.62; 33.07 -> 35.02 | Helped |
| use render-aware initialization instead of Xavier initialization (ablation) | RAKD student · Bistro, Zero-Day, Classroom, Dining room | PSNR: 25.22 -> 25.36; 26.67 -> 26.95; 29.85 -> 30.62; 33.80 -> 35.02 | Helped |
Limitations
- Performance declines on specular reflections, as shown by the mirror failure case.
- Only 10-frame sequences for training limit temporal stability compared with 64-frame methods.
- Relies on primary-ray auxiliary features (albedo, normal, depth).
Experimental setup
Datasets: Tungsten scenes PBRT scenes Bistro Zero-Day Classroom Dining room
Benchmarks: Bistro Zero-Day Classroom Dining room
Metrics: PSNR SSIM runtime per 1920x1080 frame
Baselines: NVIDIA OptiX AI denoiser (ONND) Weight-sharing model (WS) NPPD Intel Open Image Denoise (OIDN)
Builds on: Hinton knowledge distillation kernel prediction denoising Wasserstein GAN NPPD neural partitioning pyramids SMAPE loss
Closest papers by meaning
- High-Quality Real-Time Rendering Using Subpixel Sampling Reconstruction (AAAI 2024 · similarity 0.88)
- Reducing Spatial Fitting Error in Distillation of Denoising Diffusion Models (AAAI 2024 · similarity 0.86)
- Remix-DiT: Mixing Diffusion Transformers for Multi-Expert Denoising (NeurIPS 2024 · similarity 0.86)
- Noise Conditional Variational Score Distillation (ICML 2025 · similarity 0.86)
- Not All Parameters Matter: Masking Diffusion Models for Enhancing Generation Ability (CVPR 2025 · similarity 0.85)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.85)
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.