Research map / 3D vision / Face and shape
FlashAvatar: High-fidelity Head Avatar with Efficient Gaussian Embedding
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.
Builds a lightweight 3D Gaussian head avatar by attaching UV-sampled Gaussians to a FLAME mesh surface and learning an MLP offset conditioned on expression, reaching ~300FPS rendering from monocular video in minutes.
Problem
Reconstructing high-fidelity animatable head avatars from short monocular video while achieving real-time rendering and fast training, which prior NeRF-based and canonical-plus-deformation Gaussian methods struggle with.
Approach
Maintain a uniform 3D Gaussian field in UV space embedded on the FLAME mesh surface so Gaussians move with the mesh, and learn a small expression-conditioned MLP offset (position, rotation, scale) to capture non-surface regions and subtle details, keeping the Gaussian count at the 10K level.
Key results
- Achieves PSNR 32.33 and LPIPS 3.23e-2 versus INSTA's 30.54 and 3.76e-2 on public test videos.
- Renders at about 300 FPS at 512x512 on an RTX 3090, versus ~20 FPS for INSTA and ~10 FPS for PointAvatar and NHA.
- Reconstructs avatars in minutes with about 13K Gaussians at UV resolution 128 (304 FPS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase UV resolution from 64 to 128 (ablation) | FlashAvatar · Gaussian sampling density | PSNR: 30.35 -> 30.80 | Helped |
| Increase UV resolution from 64 to 128 (ablation) | FlashAvatar · Gaussian sampling density | LPIPS(10−2)↓: 4.47 -> 3.47 | Helped |
| Increase UV resolution from 64 to 128 (ablation) | FlashAvatar · Gaussian sampling density | FPS: 394 -> 304 | Hurt |
| Increase UV resolution from 128 to 256 (ablation) | FlashAvatar · Gaussian sampling density | PSNR: 30.80 -> 31.07 | Helped |
| Increase UV resolution from 128 to 256 (ablation) | FlashAvatar · Gaussian sampling density | LPIPS(10−2)↓: 3.47 -> 2.99 | Helped |
| Increase UV resolution from 128 to 256 (ablation) | FlashAvatar · Gaussian sampling density | FPS: 304 -> 112 | Hurt |
| Use uniform UV initialization instead of vertex initialization (ablation) | FlashAvatar · Head-avatar reconstruction | visual quality | Helped |
| Add a dynamic offset conditioned on expression code instead of a static offset (ablation) | FlashAvatar · Facial expression modeling | visual fidelity | Helped |
| Use mesh-attached Gaussians with learned offsets instead of a canonical Gaussian field plus deformation field (ablation) | FlashAvatar · Complex expressions and facial details | visual fidelity | Helped |
| Initialize canonical Gaussians randomly in a ball rather than on the mesh surface (ablation) | Canonical + deformation strategy · Head-avatar reconstruction | visual quality | Hurt |
Surprising findings
- Attaching Gaussians to mesh vertices fails to converge to satisfactory results, while UV-space initialization is more uniform.
- Static offset fails to model facial alterations; a dynamic expression-conditioned offset is needed.
Limitations
- Relies on good surface-embedded initialization, so tracking errors (especially global pose) cause detail loss or misalignment.
- Conditioned on tracked expression codes, so it cannot model dynamically changing hair with heavy non-rigid deformation.
- Evaluated only on monocular videos of 1-3 minutes from prior datasets.
Experimental setup
Datasets: monocular head videos released by NHA, NeRFBlendshape, PointAvatar, INSTA
Benchmarks: public monocular head avatar videos from prior works
Metrics: MSE L1 PSNR SSIM LPIPS FPS
Baselines: NHA PointAvatar INSTA Deformable 3DGS 4D-GS
Builds on: 3D Gaussian Splatting FLAME INSTA LPIPS MICA
Closest papers by meaning
- GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections (CVPR 2025 · similarity 0.92)
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.90)
- RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars (CVPR 2025 · similarity 0.90)
- GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians (CVPR 2024 · similarity 0.89)
- GASP: Gaussian Avatars with Synthetic Priors (CVPR 2025 · similarity 0.89)
- HRAvatar: High-Quality and Relightable Gaussian Head Avatar (CVPR 2025 · 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.