Research map / 3D vision / 3D generation
R-Cyclic Diffuser: Reductive and Cyclic Latent Diffusion for 3D Clothed Human Digitalization
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 R-Cyclic Diffuser, which fine-tunes Zero-1-to-3-style latent diffusion models with cyclic sequential conditioning, lateral inversion constraint, cyclic noise selection and SMPL-X blendweight priors, then fuses generated views with a multi-view pixel-aligned implicit model for single-view clothed human reconstruction.
Problem
Single-view 3D clothed human reconstruction via Zero-1-to-3-style latent diffusion suffers from view inconsistency among generated novel views and falls short of pixel-aligned implicit SOTA methods.
Approach
Replace Zero-1-to-3's independent viewpoint-conditioned generation with sequential 90-degree azimuth generation conditioned on previous predicted views, restrict outputs to an "ideal" subset of weak-perspective views, flip-align views 180 degrees apart, select noise seeds via cycle consistency, and inject SMPL-X blendweight images as conditioning.
Key results
- On THuman2.0, the SMPL-X-free variant reports CD 5.410e-4 and P2S 4.879e-4, beating IntegratedPIFu (5.925, 5.777) and Zero-1-to-3 w pixel-aligned implicit model (6.560, 5.093).
- With SMPL-X priors, the model reports CD 0.9409e-4 on THuman2.0 and 0.7119e-3 on BUFF, outperforming ICON, ECON and D-IF, which all use SMPL-X.
- Ablations on THuman2.0 show Ideal Subset Transformation raises PSNR from 14.59 to 19.74, Sequential Conditioning raises it to 20.05, and Lateral Inversion Constraint raises it to 21.24.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Ideal Subset Transformation (IST) (ablation) | Zero-1-to-3 · THuman2.0 | PSNR: 14.59 -> 19.74 | Helped |
| add Sequential Conditioning (SC) (ablation) | R-Cyclic Diffuser · THuman2.0 | PSNR: 19.74 -> 20.05 | Helped |
| add Lateral Inversion Constraint (LIC) (ablation) | R-Cyclic Diffuser · THuman2.0 | PSNR: 20.05 -> 21.24 | Helped |
| add SMPL-X Blendweight Priors (ablation) | R-Cyclic Diffuser · THuman2.0 | CD (10-4): 5.410 -> 0.9409 | Helped |
Surprising findings
- Plain Zero-1-to-3 with a pixel-aligned implicit model falls behind IntegratedPIFu on CD on THuman2.0, despite being trained on clothed human data.
Limitations
- Relies on ground-truth SMPL-X priors for the SMPL-X variant, which are not always available.
- Evaluated on a single small training set (THuman2.0 with 526 scans) and a 101-mesh BUFF subset.
Experimental setup
Datasets: THuman2.0 BUFF LAION-5B-style Internet-scale image data (as used by Stable Diffusion)
Benchmarks: THuman2.0 BUFF
Metrics: Chamfer Distance Point-to-Surface PSNR SSIM LPIPS
Baselines: Zero-1-to-3 One-2-3-45 IntegratedPIFu ICON ECON D-IF
Builds on: Zero-1-to-3 PIFu One-2-3-45 DDIM Stable Diffusion
Closest papers by meaning
- Diffusion-FOF: Single-View Clothed Human Reconstruction via Diffusion-Based Fourier Occupancy Field (CVPR 2024 · similarity 0.89)
- DeClotH: Decomposable 3D Cloth and Human Body Reconstruction from a Single Image (CVPR 2025 · similarity 0.88)
- DiffHuman: Probabilistic Photorealistic 3D Reconstruction of Humans (CVPR 2024 · similarity 0.88)
- SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion (CVPR 2024 · similarity 0.87)
- Human-3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models (NeurIPS 2024 · similarity 0.87)
- PSHuman: Photorealistic Single-image 3D Human Reconstruction using Cross-Scale Multiview Diffusion and Explicit Remeshing (CVPR 2025 · 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.