Research map / 3D vision / NeRF and radiance fields
Learning Compositional Radiance Fields of Dynamic Human Heads
Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, Michael Zollhofer
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the compositional local-animation-code radiance field rather than Neural Volumes | Dynamic human head rendering · Sequence1–Sequence4 | MSE: 46.19 -> 34.01; 52.11 -> 42.65; 83.07 -> 79.29; 40.47 -> 27.62 | Improved |
| Use the compositional local-animation-code radiance field rather than global-code NeRF | Dynamic human head rendering · Sequence1–Sequence4 | MSE: 43.34 -> 34.01; 46.89 -> 42.65; 90.45 -> 79.29; 35.52 -> 27.62 | Improved |
| Condition the radiance field on spatially varying local animation codes instead of only a global animation code (ablation) | Ours and temporally conditioned NeRF · training sequences of 1, 40, 120, 240, and 360 frames | MSE | Improved |
| Increase the coarse voxel feature resolution (ablation) | Ours · Sequence-length reconstruction evaluation | MSE | Improved |
| Reduce the coarse voxel feature resolution (ablation) | Ours · Sequence-length reconstruction evaluation | MSE | Hurt |
| Replace hierarchical sampling with simple sampling (ablation) | Ours · single sentence | Runtime: 14.30s -> 3.6s | Improved |
| Replace hierarchical sampling with simple sampling (ablation) | Ours · single sentence | SSIM: 0.9090 -> 0.9113 | No clear effect |
| Replace hierarchical sampling with simple sampling (ablation) | NeRF · single sentence | Runtime: >25s -> 19.69s | Improved |
| Replace hierarchical sampling with simple sampling (ablation) | NeRF · single sentence | SSIM: 0.8898 -> 0.8886 | No clear effect |
| Use image input rather than sparse 2D keypoints, without encoder fine-tuning (ablation) | Novel content synthesis encoder · novel content | MSE: 58.52 -> 55.86 | Improved |
| Fine-tune the encoder (ablation) | Keypoints-driven novel content synthesis · novel content | MSE: 58.52 -> 35.12 | Improved |
| Fine-tune the image encoder (ablation) | Image-driven novel content synthesis · novel content | MSE: 55.86 -> 34.86 | Improved |
Closest papers by meaning
- Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction (CVPR 2021 · similarity 0.92)
- Learning Neural Volumetric Representations of Dynamic Humans in Minutes (CVPR 2023 · similarity 0.90)
- H-NeRF: Neural Radiance Fields for Rendering and Temporal Reconstruction of Humans in Motion (NeurIPS 2021 · similarity 0.90)
- High-Fidelity 3D Head Avatars Reconstruction through Spatially-Varying Expression Conditioned Neural Radiance Field (AAAI 2024 · similarity 0.88)
- DiffPortrait360: Consistent Portrait Diffusion for 360 View Synthesis (CVPR 2025 · similarity 0.87)
- Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting (ICLR 2024 · 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.
