Research map / Image generation / Fast diffusion sampling
High-Quality Real-Time Rendering Using Subpixel Sampling Reconstruction
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 a 1/4-spp subpixel sampling strategy with a multi-scale U-Net denoiser (SSR) using temporal feature accumulation to reconstruct high-quality frames at 130 FPS at 2K resolution.
Problem
Real-time high-resolution Monte Carlo rendering is limited by sampling cost and denoiser inference time; existing denoisers target more than 1-spp and are too slow at high resolution.
Approach
Trace only one pixel per 2x2 tile per frame (1/4-spp, rotating positions over four frames), fill missing pixels via a temporal feature accumulator that blends warped history with current GBuffers using learned confidence weights, and reconstruct with a lightweight multi-scale U-Net that feeds back the warped previous output.
Key results
- SSR averages 33.10 PSNR / 0.9421 SSIM across six scenes, vs. 28.81 PSNR for NSRR (second best).
- SSR runs at 131 FPS at 1024x2048 (7.6 ms) and 220 FPS at 1024x1080 on an A100, about 37% faster than RAE.
- Subpixel sampling reduces sampling time from 12.79 ms to 4.35 ms, giving overall rendering of 6.92 ms vs 15.36 ms (about 3x).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSR instead of the strongest baseline, NSRR | Rendering denoising · Six-scene average | PSNR: 28.81 -> 33.10 | Helped |
| Use SSR instead of the strongest baseline, NSRR | Rendering denoising · Six-scene average | SSIM: .8857 -> .9421 | Helped |
| Use SSR instead of RAE | Rendering reconstruction · 1024×2048 | FPS: 96 -> 131 | Helped |
| Use SSR instead of RAE | Rendering reconstruction · 1024×1080 | FPS: 160 -> 220 | Helped |
| Use 1/4-spp subpixel sampling instead of 1-spp | Rendering pipeline · Six-scene average rendering | overall rendering time (ms): 15.36 -> 6.92 | Helped |
| Add the temporal feature accumulator (ablation) | SSR · Six-scene average | PSNR: 29.55 -> 32.13 | Helped |
| Add the temporal feature accumulator (ablation) | SSR · Six-scene average | SSIM: .8952 -> .9245 | Helped |
| Add the warped previous output (ablation) | SSR with temporal feature accumulator · Six-scene average | PSNR: 32.13 -> 33.10 | Helped |
| Add the warped previous output (ablation) | SSR with temporal feature accumulator · Six-scene average | SSIM: .9245 -> .9421 | Helped |
| Add the shadow feature (ablation) | SSR · BistroInterior | PSNR: 27.67 -> 28.22 | Helped |
| Add the transparency feature to the shadow variant (ablation) | SSR · BistroInterior | PSNR: 28.22 -> 28.99 | Helped |
| Accumulate the transparent feature before reconstruction (ablation) | SSR · Scenes with transparent features | reconstruction quality | Helped |
Surprising findings
- Predicting two coarse-scale images at the first decoder stages instead of dense features both speeds inference and improves metrics, per the authors.
Limitations
- Inference time still has room for improvement; planned 16-bit TensorRT acceleration not yet done.
- Separate model trained per scene rather than a single generalizable model.
- Swin Transformer integration improves PSNR by about 0.23 but adds 1.1 ms inference time.
Experimental setup
Datasets: BistroInterior BistroExterior Sponza Diningroom Angel Warmroom
Benchmarks: six custom rendered scenes (BistroInterior, BistroExterior, Sponza, Diningroom, Warmroom, Angel)
Metrics: PSNR SSIM RMSE FPS inference time
Baselines: AFS ANF NSRR RAE
Builds on: U-Net NSRR Chaitanya recurrent denoiser temporal anti-aliasing
Closest papers by meaning
- Real-Time Neural Denoising with Render-Aware Knowledge Distillation (AAAI 2025 · similarity 0.88)
- Low-Latency Space-Time Supersampling for Real-Time Rendering (AAAI 2024 · similarity 0.88)
- Realistic Noise Synthesis with Diffusion Models (AAAI 2025 · similarity 0.86)
- HIFA: High-fidelity Text-to-3D Generation with Advanced Diffusion Guidance (ICLR 2024 · similarity 0.85)
- ReFrame: Layer Caching for Accelerated Inference in Real-Time Rendering (ICML 2025 · similarity 0.85)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · 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.