Research map / 3D vision / Gaussian splatting
GASP: Gaussian Avatars with Synthetic Priors
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.
Trains an auto-decoder prior of mesh-attached Gaussian avatars on 1000 synthetic identities, then fits it to a single image or monocular video in three stages to get real-time 360° animatable avatars.
Problem
Creating photorealistic, real-time animatable avatars with free-viewpoint (360°) rendering from only a single image or monocular webcam/smartphone video, where unseen regions (back/sides of head) are ill-posed without a prior.
Approach
Train a generative prior over Gaussian avatar parameters on large-scale synthetic data (per-Gaussian learnable features plus identity latent, canonical template offsets), then fit to real users with a three-stage process (latent inversion, MLP fine-tuning, Gaussian refinement) and a prior regularization term so that learned feature correlations transfer unseen-region edits from the front to the ba
Key results
- In the monocular video setting, GASP reaches PSNR 21.34, SSIM 0.712, LPIPS 0.333, FID 117 and user QUAL 3.68, beating FlashAvatar (17.25 PSNR) and GaussianAvatars (17.39 PSNR).
- In the single image setting, GASP achieves PSNR 20.73 and FID 119, outperforming ROME on all metrics.
- Avatars render at 70fps on an NVIDIA 4090 with no neural networks at inference and about 15MB per avatar.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GASP instead of DiffusionRig | Monocular avatar fitting · Monocular video | PSNR: 19.67 -> 21.34 | Helped |
| Use GASP instead of HeadNeRF | Single-image avatar fitting · Single image | PSNR: 17.72 -> 20.73 | Helped |
| Use GASP instead of FlashAvatar | Multi-camera avatar fitting · Multi-camera | FID: 125 -> 101 | Helped |
| Add the synthetic prior (ablation) | GASP monocular fitting · Monocular video | PSNR: 19.42 -> 20.67 | Helped |
| Add prior regularization (ablation) | GASP monocular fitting · Monocular video | PSNR: 20.31 -> 20.67 | Helped |
| Add fitting stage 1 (optimize identity latent z) (ablation) | GASP monocular fitting · Monocular video | PSNR: 19.56 -> 20.67 | Helped |
| Add fitting stage 2 (fine-tune decoder) (ablation) | GASP monocular fitting · Monocular video | PSNR: 20.33 -> 20.67 | Helped |
| Add fitting stage 3 (refine Gaussians) (ablation) | GASP monocular fitting · Monocular video | ID-SIM: 0.441 -> 0.589 | Helped |
| Increase prior training identities from 1 to 1k (ablation) | GASP prior · Monocular video | PSNR: 15.86 -> 20.67 | Helped |
| Increase prior training identities from 10 to 100 (ablation) | GASP prior · Monocular video | PSNR: 19.98 -> 20.39 | Helped |
| Increase Gaussian map resolution from 64×64 to 512×512 (ablation) | GASP avatar · Monocular video | PSNR: 20.41 -> 20.67 | Helped |
| Remove prior regularization (ablation) | GASP monocular fitting · Monocular video | ID-SIM: 0.589 -> 0.620 | Helped |
Surprising findings
- Using a prior trained on only one synthetic subject performs worse than no prior at all (PSNR 15.86 vs 19.42 without prior).
- Freezing per-Gaussian features during fitting lets edits at the front of the head propagate to unseen back-of-head Gaussians via learned semantic correlations.
Limitations
- Back of the head looks synthetic-looking since prior is trained on synthetic data.
- Only flat RGB colour without lighting modeling, reducing realism.
- Underperforms other Gaussian Avatar methods in the multi-camera setting, as the prior regularizes toward synthetic solutions.
Experimental setup
Datasets: synthetic human dataset (Hewitt et al. pipeline, 1000 identities) NeRSemble
Benchmarks: NeRSemble
Metrics: PSNR SSIM LPIPS FID ID-SIM user rating QUAL
Baselines: FlashAvatar GaussianAvatars ROME HeadNeRF FLARE DiffusionRig
Builds on: 3D Gaussian Splatting FlashAvatar GaussianAvatars auto-decoder DeepSDF-style priors Cafca
Closest papers by meaning
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.92)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.92)
- ASH: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering (CVPR 2024 · similarity 0.91)
- GAvatar: Animatable 3D Gaussian Avatars with Implicit Mesh Learning (CVPR 2024 · similarity 0.91)
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.90)
- GAF: Gaussian Avatar Reconstruction from Monocular Videos via Multi-view Diffusion (CVPR 2025 · similarity 0.90)
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.