Research map / 3D vision / Gaussian splatting
HumanGaussian: Text-Driven 3D Human Generation with Gaussian Splatting
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 HumanGaussian, a 3D Gaussian Splatting framework for text-to-3D human generation that uses SMPL-X-initialized Gaussians, a joint RGB-depth diffusion prior, annealed negative-prompt SDS, and size-based pruning.
Problem
Text-driven 3D human generation via score distillation struggles to balance fine geometric detail, realistic appearance (over-saturation), and training efficiency across mesh, NeRF, and existing 3DGS approaches.
Approach
Initialize 3D Gaussians densely on the SMPL-X surface, distill jointly from an RGB-and-depth Stable Diffusion variant conditioned on pose skeletons, decompose SDS into classifier and annealed negative-prompt scores to avoid over-saturation at nominal CFG, and prune large-scale Gaussians to remove floaters.
Key results
- User study MOS: HumanGaussian scored 4.24 texture, 3.88 geometry, 4.71 text alignment, the best of all compared methods (TADA 3.76/3.53/4.35).
- Frontal-view metrics: CLIP 30.82, aesthetic 6.436, HPSv2 0.262, all highest among compared methods (TADA CLIP 30.13, aesthetic 5.709).
- Full framework trains in about one hour on a single NVIDIA A100 (40GB) GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HumanGaussian instead of TADA | 3D human generation · 3D human generation, user study | Texture Quality: 3.76 -> 4.24 | Helped |
| Use HumanGaussian instead of DreamHuman | 3D human generation · 3D human generation, user study | Geometry Quality: 3.65 -> 3.88 | Helped |
| Use HumanGaussian instead of TADA | 3D human generation · 3D human generation, user study | Text Alignment: 4.35 -> 4.71 | Helped |
| Use HumanGaussian instead of TADA | 3D human generation · 3D human generation, frontal views | CLIP: 30.13 -> 30.82 | Helped |
| Use HumanGaussian instead of TADA | 3D human generation · 3D human generation, frontal views | Aesthetic score: 5.709 -> 6.436 | Helped |
| Use HumanGaussian instead of TADA | 3D human generation · 3D human generation, frontal views | HPSv2: 0.253 -> 0.262 | Helped |
| Add SMPL-X initialization and pose conditioning (ablation) | HumanGaussian · Text-driven 3D human generation | body structure coherence | Helped |
| Add annealed negative prompt guidance (ablation) | HumanGaussian · Text-driven 3D human generation | appearance realism | Helped |
| Add depth-branch SDS (ablation) | HumanGaussian · Text-driven 3D human generation | geometric errors | Helped |
| Add size-conditioned Gaussian pruning (ablation) | HumanGaussian · Text-driven 3D human generation | floating artifacts | Helped |
Limitations
- Existing T2I models perform poorly on hands and feet, so these parts are sometimes rendered unfaithfully.
- Size-based pruning may mistakenly remove some useful Gaussians.
- DreamHuman results were taken from its paper and project page because no official code or model was released.
Experimental setup
Datasets: LAION MiDaS depth labels
Benchmarks: user study (30 DreamHuman prompts) CLIP score aesthetic score HPSv2
Metrics: MOS texture quality MOS geometry quality MOS text alignment CLIP score aesthetic score HPSv2
Baselines: TADA DreamHuman DreamGaussian GaussianDreamer HumanNorm
Builds on: DreamFusion 3D Gaussian Splatting SMPL-X ThreeStudio Stable Diffusion
Closest papers by meaning
- GaussianIP: Identity-Preserving Realistic 3D Human Generation via Human-Centric Diffusion Prior (CVPR 2025 · similarity 0.92)
- PaintHuman: Towards High-Fidelity Text-to-3D Human Texturing via Denoised Score Distillation (AAAI 2024 · similarity 0.89)
- GauHuman: Articulated Gaussian Splatting from Monocular Human Videos (CVPR 2024 · similarity 0.89)
- DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation (ICLR 2024 · similarity 0.89)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.88)
- Text-to-3D using Gaussian Splatting (CVPR 2024 · similarity 0.88)
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.