Research map / 3D vision / Human pose estimation
IDOL: Instant Photorealistic 3D Human Creation from a Single Image
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.
Introduces HuGe100K, a 100K-subject generated multi-view human dataset, and IDOL, a feed-forward transformer that predicts animatable 3D Gaussian human avatars in SMPL-X UV space from a single image in about a second.
Problem
Single-image, animatable, photorealistic 3D full-body human reconstruction is limited by scarce high-quality multi-view human training data and by slow, fragile optimization-based or non-animatable methods.
Approach
Scale up training data by generating multi-view human videos with a fine-tuned pose-controllable video model (MVChamp, built on Champ) from FLUX-generated images, then train a feed-forward transformer (Sapiens encoder + UV-alignment transformer) that predicts SMPL-X UV-space Gaussian attribute maps for direct animation.
Key results
- IDOL-full achieves MSE 0.008, PSNR 21.673, LPIPS 1.138 versus SIFU (PSNR 14.204) and GTA (PSNR 14.282) on the held-out set.
- Removing HuGe100K training (Ours-w/o HuGe100K) drops PSNR to 19.225, showing the value of the generated data.
- The model reconstructs photorealistic 1K-resolution avatars from one image in under a second on a single A100 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HuGe100K to training (ablation) | IDOL · Human reconstruction test set | MSE: 0.017 -> 0.008 | Helped |
| Add HuGe100K to training (ablation) | IDOL · Human reconstruction test set | PSNR: 19.225 -> 21.673 | Helped |
| Add HuGe100K to training (ablation) | IDOL · Human reconstruction test set | LPIPS: 1.326 -> 1.138 | Helped |
| Use the Sapiens encoder instead of DINOv2 (ablation) | IDOL · Human reconstruction | Texture and fold detail quality | Helped |
Surprising findings
- Optimization-based SIFU and GTA score lower than expected partly because orthographic projection assumptions mismatch the perspective test cameras, per the authors.
Limitations
- Video model limits synthesis to single-frame images from fixed viewpoints, with no long motion sequences.
- Facial identity and expression are not specifically modeled, and facial optimization is secondary.
- Handling half-body inputs remains challenging.
Experimental setup
Datasets: HuGe100K THuman 2.1 DeepFashion
Benchmarks: HuGe100K test split THuman 2.1 test split
Metrics: MSE PSNR LPIPS
Baselines: GTA SIFU LGM DreamGaussian Ours-w/o HuGe100K
Builds on: Champ 3D Gaussian Splatting E3Gen Sapiens LGM
Closest papers by meaning
- SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors (CVPR 2025 · similarity 0.88)
- AniGS: Animatable Gaussian Avatar from a Single Image with Inconsistent Gaussian Reconstruction (CVPR 2025 · similarity 0.88)
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting (CVPR 2024 · similarity 0.88)
- Human Gaussian Splatting: Real-time Rendering of Animatable Avatars (CVPR 2024 · similarity 0.88)
- Real-time High-fidelity Gaussian Human Avatars with Position-based Interpolation of Spatially Distributed MLPs (CVPR 2025 · similarity 0.88)
- 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.