Research map / Architectures and efficiency / RNNs and normalization
Achieving Lightweight Super-Resolution for Real-Time Computer Graphics
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 CGSR, a framework that injects real-time rendering info (depth, normals, object edges) into lightweight SR backbones, then applies evolutionary NAS and rendering-aware dynamic pruning to cut parameters and ops.
Problem
Learning-based super-resolution for real-time computer graphics is too computationally heavy, undermining the rendering savings from low-resolution rendering.
Approach
Use readily available rendering buffers (depth, normal directions, object edges) as intermediate features inserted between feature extraction and reconstruction of an existing SR network, optimize the result with rendering-guided evolutionary NAS, and prune weights dynamically using temporal rendering patterns.
Key results
- CGSR-P (optimized PAN) reduces parameters by about 77% and multi-adds by about 78% versus PAN while keeping PSNR/SSIM comparable (e.g., Doom3 x2: 62k params, 15G M-A, 36.11 dB vs PAN 261k, 71G, 36.16 dB).
- Across backbones (SRCNN, EDSR, PAN), CGSR reduces inference latency by an average of 30% with similar quality.
- Rendering-aware hybrid pruning gives a further 33% parameter reduction over magnitude-based pruning, with under 0.05 dB PSNR loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize PAN with CG-guided NAS and rendering-aware hybrid pruning | PAN · Doom3 ×2 | PSNR: 36.16 -> 36.11 | No clear effect |
| Optimize PAN with CG-guided NAS and rendering-aware hybrid pruning | PAN · Quake 4 ×2 | PSNR: 33.15 -> 33.07 | No clear effect |
| Optimize PAN with CG-guided NAS and rendering-aware hybrid pruning | PAN · The Chronicles of Riddick ×2 | PSNR: 38.01 -> 38.16 | Helped |
| Optimize PAN with CG-guided NAS and rendering-aware hybrid pruning | PAN · Doom3 ×3 | PSNR: 34.62 -> 34.55 | No clear effect |
| Optimize PAN with CG-guided NAS and rendering-aware hybrid pruning | PAN · Quake 4 ×3 | PSNR: 30.18 -> 30.20 | No clear effect |
| Optimize PAN with CG-guided NAS and rendering-aware hybrid pruning | PAN · The Chronicles of Riddick ×3 | PSNR: 31.53 -> 31.51 | No clear effect |
| Add depth rendering information (ablation) | CGSR-P · Doom3 trace | parameter size | Helped |
| Add normal-direction rendering information (ablation) | CGSR-P · Doom3 trace | SR quality | Helped |
| Add object-edge rendering information (ablation) | CGSR-P · Doom3 trace | SR quality | Helped |
| Add depth, normal-direction, and object-edge rendering information together (ablation) | CGSR-P · Doom3 trace | SR quality: 37% | Helped |
| Use CG-guided NAS to shrink the PAN network (ablation) | PAN · Doom3 trace | parameter size: 37% | Helped |
| Use rendering-aware hybrid pruning instead of magnitude-based pruning (ablation) | CG-optimized PAN · Doom3 trace | parameter size: 33% | Helped |
Surprising findings
- Adding rendering information reduces parameter count while the SR quality is maintained or improved, which the authors attribute to the rendering features helping the network retain learned knowledge.
Limitations
- Hardware scheduling may cause workload balance issues for architecture shrinking and dynamic pruning.
- Rendering data comes from an RTL-level GPU simulator, limiting rendering information types and possibly introducing errors in training data.
- Evaluated only on three game traces and CNN-based SR backbones.
Experimental setup
Datasets: Doom 3 game trace Quake 4 game trace The Chronicles of Riddick game trace
Benchmarks: Doom 3 trace Quake 4 trace tCoR trace
Metrics: PSNR SSIM parameter count multi-add operations inference latency
Baselines: SRCNN EDSR PAN FSR VDSR magnitude-based pruning
Builds on: PAN EDSR SRCNN Neural Supersampling
Closest papers by meaning
- Efficient Video Super-Resolution for Real-time Rendering with Decoupled G-buffer Guidance (CVPR 2025 · similarity 0.89)
- Lightweight Image Super-Resolution via Flexible Meta Pruning (ICML 2024 · similarity 0.87)
- Neural Super-Resolution for Real-time Rendering with Radiance Demodulation (CVPR 2024 · similarity 0.86)
- Low-Latency Space-Time Supersampling for Real-Time Rendering (AAAI 2024 · similarity 0.86)
- Image Processing GNN: Breaking Rigidity in Super-Resolution (CVPR 2024 · similarity 0.86)
- High-Quality Real-Time Rendering Using Subpixel Sampling Reconstruction (AAAI 2024 · similarity 0.84)
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.