Research map / 3D vision / Gaussian splatting
LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh
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.
Proposes LIFe-GoM, a feed-forward generalizable human renderer that combines iterative feedback updates with a coupled multi-resolution Gaussians-on-Mesh representation to render animatable avatars from sparse views in under 1s reconstruction at 95 FPS.
Problem
Generalizable animatable human rendering from sparse inputs must reconstruct the canonical human representation in a single forward pass, while keeping rendering efficient and high-resolution.
Approach
Iteratively refine the canonical Gaussians-on-Mesh representation with feedback from rendered source views inside a feed-forward network, and decouple a low-resolution mesh (for reconstruction and articulation) from high-resolution Gaussians attached to a subdivided mesh (for rendering).
Key results
- On THuman2.0 with 3 source views, LIFe-GoM reaches 24.65 PSNR / 110.82 LPIPS* / 51.27 FID, versus GHG's 21.90 / 133.41 / 61.67.
- Rendering runs at 95.1 FPS at 1024x1024 on an A100 with reconstruction of about 908 ms at T=3 iterations.
- On AIST++, LIFe-GoM achieves 25.25 PSNR / 0.9812 SSIM / 21.61 LPIPS* in 589 ms reconstruction, comparable to ActorsNeRF (25.23 / 0.9809 / 22.45).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add iterative feedback updates (3 iterations rather than 1) (ablation) | LIFe-GoM · THuman2.0, 3 views | PSNR: 23.74 -> 24.65 | Helped |
| increase feedback iterations from 3 to 4 or 5 (ablation) | LIFe-GoM · THuman2.0, 3 views | PSNR: 24.65 -> 24.69; 24.65 -> 24.70 | Helped |
| increase feedback iterations from 3 to 4 or 5 (ablation) | LIFe-GoM · THuman2.0, 3 views | Reconstruction time: 907.92 -> 1198.14; 907.92 -> 1563.92 | Hurt |
| subdivide the mesh once or twice to attach Gaussians to a higher-resolution mesh (ablation) | LIFe-GoM · THuman2.0, 3 views | LPIPS*: 140.60 -> 118.64 -> 110.82 | Helped |
| subdivide the mesh twice rather than not subdividing (ablation) | LIFe-GoM · THuman2.0, 3 views | PSNR: 24.76 -> 24.65 | Mixed |
| update low-resolution mesh vertices rather than high-resolution mesh vertices (ablation) | LIFe-GoM · THuman2.0, 3 views | Reconstruction time: 12.45s -> 907.92ms | Helped |
| use SMPL-X shape input for mesh initialization (ablation) | LIFe-GoM · THuman2.0, 3 views | PSNR: 24.15 -> 24.65 | Helped |
| use LIFe-GoM rather than GHG | generalizable human rendering · THuman2.0, 3 views | PSNR: 21.90 -> 24.65 | Helped |
| use LIFe-GoM rather than GHG | cross-domain generalization · XHuman | PSNR: 23.52 -> 25.32 | Helped |
| use LIFe-GoM rather than ActorsNeRF | generalizable human rendering · AIST++ | LPIPS*: 22.45 -> 21.61 | Helped |
| add Gaussian pose noise at increasing standard deviations (ablation) | LIFe-GoM · THuman2.0, noisy input poses | PSNR: 23.96 -> 22.02 -> 20.43 | Hurt |
| use LIFe-GoM rather than GHG under noisy input poses | generalizable human rendering · THuman2.0, pose noise std=0.5 | PSNR: 18.53 -> 20.43 | Helped |
Surprising findings
- Without SMPL-X shape input (average shape initialization) performance drops only slightly (24.15 vs 24.65 PSNR), still beating GHG with shape input.
- PSNR did not consistently improve with more subdivisions, as the authors note PSNR sometimes prefers blurry results.
Limitations
- Cannot inpaint large invisible regions in source images without a hallucination module.
- Mesh topology fixed to SMPL/SMPL-X, so clothing like dresses and coats is not handled.
- Fails on unseen clothing types and is sensitive to input pose accuracy (noise degrades results).
Experimental setup
Datasets: THuman2.0 AIST++ XHuman
Benchmarks: THuman2.0 XHuman AIST++
Metrics: PSNR LPIPS* FID SSIM
Baselines: GoMAvatar 3DGS-Avatar iHuman NHP NIA GHG GPS-Gaussian HumanNeRF ActorsNeRF
Builds on: GoMAvatar Gaussian splatting Point Transformer GHG
Closest papers by meaning
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian Reconstruction (CVPR 2025 · similarity 0.88)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.88)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.87)
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.87)
- GoMAvatar: Efficient Animatable Human Modeling from Monocular Video Using Gaussians-on-Mesh (CVPR 2024 · similarity 0.87)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 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.