Research map / 3D vision / 3D generation
SIGMAN: Scaling 3D Human Gaussian Generation with Millions of Assets
Yuhang Yang, Fengqi Liu, Yixing Lu, Qin Zhao, Pingyu Wu, Wei Zhai, Ran Yi, Yang Cao, Lizhuang Ma, Zheng-Jun Zha, Junting Dong
Read the paperCheck 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 UV-structured latent tokens vs explicit SMPL-X point prior in VAE (ablation) | SIGMAN VAE · HGS-1M VAE reconstruction | PSNR: 28.61 -> 29.32 | Improved |
| add Conv-Atten dual branch (ablation) | SIGMAN VAE · HGS-1M VAE ablation | PSNR: 29.04 -> 29.32 | Improved |
| add UV-map initialization of learnable tokens (ablation) | SIGMAN VAE · HGS-1M VAE ablation | PSNR: 28.92 -> 29.32 | Improved |
| increase VAE latent channels from 16 to 32 (ablation) | SIGMAN VAE · HGS-1M VAE ablation | PSNR: 29.32 -> 29.78 (VAE); generation harder | Mixed |
| increase VAE latent channels from 4 to 16 (ablation) | SIGMAN VAE · HGS-1M VAE ablation | PSNR: 25.14 -> 29.32 | Improved |
| scale the VAE up from 146M (ablation) | SIGMAN VAE · HGS-1M VAE ablation | PSNR: 29.14 -> 29.32 | Improved |
| use MM-DiT instead of Cross-DiT (ablation) | SIGMAN generation · HGS-1M image-to-3D | PSNR: 24.72 -> 25.45 | Improved |
| use Sapiens human encoder instead of CLIP or DINOv2 (ablation) | SIGMAN generation · HGS-1M image-to-3D | PSNR: 24.07 -> 25.45 (CLIP); 24.95 -> 25.45 (DINOv2) | Improved |
| use rectified flow instead of DDPM v-prediction (ablation) | SIGMAN generation · HGS-1M image-to-3D | PSNR: 25.45 vs 25.38 | No clear effect |
| scale DiT from 200M to 2B (ablation) | SIGMAN generation · HGS-1M image-to-3D | PSNR: 23.73 -> 24.70 -> 25.45 | Improved |
| train on the whole HGS-1M dataset vs 300k subset | SIGMAN · HGS-1M image-to-3D | PSNR: 25.45 -> 26.14 | Improved |
| latent generation with UV VAE and MM-DiT vs optimization, cascade and regression baselines | SIGMAN · HGS-1M image-to-3D | PSNR: 21.03 -> 25.45 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DiffSplat — DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation (ICLR 2025) · better on 1 of 1 shared result
- SIFU — SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- HumanGaussian: Text-Driven 3D Human Generation with Gaussian Splatting (CVPR 2024 · similarity 0.89)
- GaussianIP: Identity-Preserving Realistic 3D Human Generation via Human-Centric Diffusion Prior (CVPR 2025 · similarity 0.89)
- Gaussian Shell Maps for Efficient 3D Human Generation (CVPR 2024 · similarity 0.88)
- SyncHuman: Synchronizing 2D and 3D Generative Models for Single-view Human Reconstruction (NeurIPS 2025 · similarity 0.87)
- Generalizable Human Gaussians from Single-View Image (ICLR 2025 · similarity 0.87)
- HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion (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.
