Research map / 3D vision / Gaussian splatting
HiFi4G: High-Fidelity Human Performance Rendering via Compact Gaussian Splatting
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.
Presents HiFi4G, a compact explicit 4D Gaussian Splatting method for high-fidelity human performance rendering that marries 3DGS with non-rigid tracking (dual-graph, temporal/smooth regularizers) plus residual-based compression to under 2 MB per frame.
Problem
Efficiently rendering realistic, high-fidelity human performances from dense multi-view video in a rasterization-friendly, compact representation, while avoiding temporal jitter and large per-frame storage.
Approach
Combine 3D Gaussian Splatting with embedded-deformation non-rigid tracking via a dual-graph (coarse ED graph for motion priors, fine Gaussian graph for per-kernel motion), regularize with temporal and adaptive-weighted smooth terms, and compress residuals relative to keyframes with quantization and RANS entropy coding.
Key results
- On the authors' dataset, HiFi4G after compression reaches PSNR 35.788 and VMAF 84.312 at 1.648 MB per frame, versus HumanRF's 31.174 PSNR and 11.38 MB.
- On ActorsHQ, HiFi4G after compression achieves PSNR 34.704 and 2.143 MB per frame, compared with HumanRF's 34.106 PSNR at 10.95 MB.
- Compression yields roughly 25x reduction (43.42 MB to 1.648 MB per frame on the own dataset) with less than 2 MB per frame storage; optimization takes about 4 minutes per frame on an RTX 3090.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the coarse ED-graph prior (ablation) | HiFi4G · Compact 4D Gaussians | PSNR: 29.142 -> 35.085 | Helped |
| add the Gaussian graph (ablation) | HiFi4G · Compact 4D Gaussians | PSNR: 31.185 -> 35.085 | Helped |
| add temporal regularization Etemp (ablation) | HiFi4G · Compact 4D Gaussians | PSNR: 33.555 -> 35.085 | Helped |
| add smooth regularization Esmooth (ablation) | HiFi4G · Compact 4D Gaussians | PSNR: 33.889 -> 35.085 | Helped |
| use adaptive weight wi,t instead of fixed weight 0.1 (ablation) | HiFi4G · Compact 4D Gaussians | PSNR: 33.577 -> 35.085 | Helped |
| apply residual compensation with low-bit quantization (ablation) | HiFi4G compression · Per-frame encoding | Per-frame storage: 3.67MB to under 2MB per frame | Helped |
Surprising findings
- Dynamic 3D Gaussians has much lower PSNR (22.413) than other methods on ActorsHQ, indicating fixed appearance attributes hurt quality on that dataset.
Limitations
- Relies heavily on foreground segmentation; poor segmentation causes artifacts in human-object interaction scenes.
- Requires per-frame reconstruction and mesh tracking; Gaussian optimization still takes several minutes per frame.
- Dependence on fast GPU sorting limits deployment on web viewers and mobile devices.
Experimental setup
Datasets: own multi-view capture (81 Z-CAM cameras) ActorsHQ
Benchmarks: own captured dataset ActorsHQ
Metrics: PSNR SSIM LPIPS VMAF per-frame storage (MB)
Baselines: Instant-NSR NeuS2 HumanRF Dynamic 3D Gaussians
Builds on: 3D Gaussian Splatting DynamicFusion Fusion4D embedded deformation NeuS2
Closest papers by meaning
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.90)
- Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis (CVPR 2024 · similarity 0.89)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.89)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- GauHuman: Articulated Gaussian Splatting from Monocular Human Videos (CVPR 2024 · similarity 0.89)
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.