Research map / 3D vision / Face and shape
Gaussian and implicit head avatars
Builds animatable head avatars from video using morphable, Gaussian blendshape or implicit representations.
115 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Face and shape, keyword cluster “avatar · human · body”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
54 papers in 2023–24 and 50 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 14.9% to 13.7%.
Open questions the data points to
- Papers report GaussianAvatars on NeRSemble novel-view synthesis (LPIPS) anywhere from 0.055 to 0.067: comparisons against it depend on whose number you take.
- Papers report GaussianAvatars on NeRSemble novel-view synthesis (PSNR) anywhere from 30.29 to 33.54: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2019 · Textured Neural Avatars (CVPR 2019 · 112 citations)
- 2021 · Pixel Codec Avatars (CVPR 2021 · 109 citations)
- 2022 · Neural Head Avatars From Monocular RGB Videos (CVPR 2022 · 197 citations)
- 2023 · Instant Volumetric Head Avatars (CVPR 2023 · 174 citations)
- 2024 · GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · 241 citations)
- 2025 · HRAvatar: High-Quality and Relightable Gaussian Head Avatar (CVPR 2025 · 16 citations)
- 2026 · PFAvatar: Pose-Fusion 3D Personalized Avatar Reconstruction from Real-World Outfit-of-the-Day Photos (AAAI 2026 · 1 citation)
Earliest papers here
- Textured Neural Avatars (CVPR 2019 · 112 citations)
- Pixel Codec Avatars (CVPR 2021 · 109 citations)
- ANR: Articulated Neural Rendering for Virtual Avatars (CVPR 2021 · 56 citations)
What has been tried
87 papers here report 836 measured changes (a component added, removed or swapped, compared with a baseline): helped 78%, no clear effect 4%, hurt 16%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add adaptive density control with binding inheritance | Novel-View Synthesis · LPIPS | 0.206 -> 0.098 | Helped | GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · 241 citations) |
| Add joint motion optimization to the baseline | People-Snapshot · PSNR | 27.71 -> 30.98; 25.09 -> 28.77; 24.80 -> 29.59; 26.16 -> 30.83 | Helped | GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians (CVPR 2024 · 171 citations) |
| Add the super-resolution network | Self reenactment · LPIPS (2K) | 0.202 -> 0.098 | Helped | Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · 109 citations) |
| Use the proposed radiance-transfer appearance instead of EyeNeRF | Held-out segments · PSNR | 35.110 -> 36.501 | Helped | Relightable Gaussian Codec Avatars (CVPR 2024 · 102 citations) |
| Add SMPL-derived density fields | 3D human avatar generation · geometry and texture quality | Helped | DreamAvatar: Text-and-Shape Guided 3D Human Avatar Generation via Diffusion Models (CVPR 2024 · 100 citations) | |
| Increase UV resolution from 64 to 128 | Gaussian sampling density · PSNR | 30.35 -> 30.80 | Helped | FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding (CVPR 2024 · 77 citations) |
| add expression augmentation | Hybrid test set · LSE-D | 11.6229 -> 11.2737 | Helped | FaceTalk: Audio-Driven Motion Diffusion for Neural Parametric Head Models (CVPR 2024 · 49 citations) |
| Use diffusion instead of single-pass output prediction | NeRSemble self-reenactment · LPIPS | 0.133 -> 0.074 | Helped | DiffusionAvatars: Deferred Diffusion for High-fidelity 3D Head Avatars (CVPR 2024 · 31 citations) |
| add disentangled learning | FFHQ reconstruction · LPIPSh | 0.244 -> 0.238 | Helped | Portrait4D: Learning One-Shot 4D Head Avatar Synthesis using Synthetic Data (CVPR 2024 · 28 citations) |
| add fine canonical distance to coarse GS-KNN distance in HDQ | SyntheticHuman++ sparse-view, jody · Relighting PSNR | 19.56 -> 21.57 | Helped | Relightable and Animatable Neural Avatar from Sparse-View Video (CVPR 2024 · 23 citations) |
| Replace volumetric scattering with rendered depth and surface scattering | RANA subject 01 · Albedo PSNR | 24.1 -> 24.05 | Hurt | IntrinsicAvatar: Physically Based Inverse Rendering of Dynamic Humans from Monocular Videos via Explicit Ray Tracing (CVPR 2024 · 22 citations) |
| add the Refiner | SHOW test · FGD (Face) | 15.49 -> 5.59 | Helped | Towards Variable and Coordinated Holistic Co-Speech Motion Generation (CVPR 2024 · 21 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| GaussianAvatars | NeRSemble novel-view synthesis | LPIPS | 0.055 – 0.067 | 2 papers, e.g. CVPR 2025 and ICLR 2025 |
| GaussianAvatars | NeRSemble novel-view synthesis | PSNR | 30.29 – 33.54 | 2 papers, e.g. ICLR 2025 and CVPR 2025 |
What reviewers will expect
Datasets and baselines named in the 27 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- NeRSemble · 3 of 27 papers
- ZJU-MoCap · 2 of 27 papers
- INSTA · 2 of 27 papers
- VFHQ · 2 of 27 papers
Baselines
- SplattingAvatar · 6 of 27 papers
- FlashAvatar · 4 of 27 papers
- GaussianAvatars · 4 of 27 papers
- INSTA · 4 of 27 papers
- PointAvatar · 4 of 27 papers
- PoseVocab · 2 of 27 papers
- HumanGaussian · 2 of 27 papers
- MonoGaussianAvatar · 2 of 27 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 104 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 8 papers
- Meta · 6 papers
- Max Planck Institute for Intelligent Systems · 5 papers
- ETH Zurich · 5 papers
- Google · 4 papers
- ByteDance · 4 papers
- Tencent · 4 papers
- Hong Kong University of Science and Technology · 4 papers
- Alibaba Group · 3 papers
- Carnegie Mellon University · 3 papers
Use it
Download BibTeX of 34 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Full-body avatar reconstruction (78 papers)
- Clothed human modeling (60 papers)
- Generative textures for 3D humans (27 papers)
Elsewhere in the topic
- Neural implicit surface reconstruction (131 papers)
- Textured mesh generation (111 papers)
- Primitive-based shape abstraction (99 papers)
- Single-view 3D reconstruction (96 papers)
- Face reflectance and relightable rendering (85 papers)
- 3D morphable face model regression (80 papers)
- Articulated and non-rigid shape learning (75 papers)
- 3D shape generative modeling (74 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).