Research map / 3D vision / Gaussian splatting
4K4D: Real-Time 4D View Synthesis at 4K Resolution
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 4K4D, a 4D point-cloud representation with a 4D feature grid, hybrid discrete image-blending plus SH appearance model, and differentiable depth peeling on hardware rasterization, enabling real-time dynamic view synthesis at 4K.
Problem
Real-time, high-fidelity novel view synthesis of dynamic 3D scenes at 4K resolution, where prior dynamic NeRF-style methods are too slow at high resolution.
Approach
Represent a dynamic scene as a semi-transparent point cloud whose per-point features come from a K-Planes 4D feature grid; make the image-blending appearance view-direction-independent so it can be pre-computed, complement it with a continuous SH term, and render via a differentiable depth peeling shader on the hardware rasterizer.
Key results
- On DNA-Rendering, 4K4D reaches PSNR 31.173, LPIPS 0.055 at 203.6 FPS vs ENeRF's 28.108 PSNR at 6.0 FPS.
- On ENeRF-Outdoor, 4K4D achieves PSNR 25.815 and LPIPS 0.147 at 141.7 FPS, versus ENeRF at 11.3 FPS.
- Rendering reaches 80-85 FPS at 4K on an RTX 4090 (ENeRF-Outdoor 85.1 FPS; DNA-Rendering 288.8 FPS with GUI), with about 2 MB per frame storage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the 4D embedding instead of per-frame, per-point optimizable parameters (ablation) | 4K4D · DNA-Rendering, 0013 01 | PSNR: 29.779 -> 31.990 | Helped |
| Add the image-blending appearance term (ablation) | 4K4D · DNA-Rendering, 0013 01 | PSNR: 30.259 -> 31.990 | Helped |
| Add the SH appearance term (ablation) | 4K4D · DNA-Rendering, 0013 01 | PSNR: 31.946 -> 31.990 | Helped |
| Add the LPIPS loss (ablation) | 4K4D · DNA-Rendering, 0013 01 | LPIPS: 0.063 -> 0.040 | Helped |
| Add the LPIPS loss (ablation) | 4K4D · DNA-Rendering, 0013 01 | PSNR: 31.661 -> 31.990 | Helped |
| Add the mask loss (ablation) | 4K4D · DNA-Rendering, 0013 01 | PSNR: 29.115 -> 31.990 | Helped |
| Add the mask loss (ablation) | 4K4D · DNA-Rendering, 0013 01 | LPIPS: 0.073 -> 0.040 | Helped |
| Increase the SH-coefficient degree (ablation) | 4K4D · DNA-Rendering | PSNR: 30.129 vs. 30.259 | No clear effect |
Surprising findings
- 3DGS overfits more on ENeRF-Outdoor with fewer views, yielding PSNR 21.63 vs 26.54 for 4K4D on the first frame.
Limitations
- Cannot produce point correspondences across frames, which downstream tasks may need.
- Storage grows linearly with the number of video frames, making long volumetric videos difficult.
- Rendering quality depends on input image resolution; 4K-quality output needs high-resolution inputs.
Experimental setup
Datasets: DNA-Rendering ENeRF-Outdoor NHR Neural3DV
Benchmarks: DNA-Rendering ENeRF-Outdoor
Metrics: PSNR SSIM LPIPS FPS storage
Baselines: ENeRF IBRNet K-Planes Im4D 3DGS
Builds on: K-Planes ENeRF 3D Gaussian Splatting Instant-NGP IBRNet
Closest papers by meaning
- 4D-LRM: Large Space-Time Reconstruction Model From and To Any View at Any Time (NeurIPS 2025 · similarity 0.89)
- EG4D: Explicit Generation of 4D Object without Score Distillation (ICLR 2025 · similarity 0.88)
- CAT4D: Create Anything in 4D with Multi-View Video Diffusion Models (CVPR 2025 · similarity 0.88)
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · similarity 0.87)
- Instant4D: 4D Gaussian Splatting in Minutes (NeurIPS 2025 · similarity 0.87)
- Robust Multi-Object 4D Generation for In-the-wild Videos (CVPR 2025 · similarity 0.87)
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.