Research map / 3D vision / Face and shape
ConTex-Human: Free-View Rendering of Human from a Single Image with Texture-Consistent Synthesis
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.
Renders free-view 3D humans from a single image by synthesizing a texture-consistent back view via depth- and text-guided attention injection, then mapping textures onto a DMTet mesh with a visibility-aware patch consistency loss.
Problem
Free-view rendering of a 3D human from one image suffers from over-smooth textures in generalizable implicit fields and texture-inconsistent, over-saturated invisible regions from SDS-based image-to-3D methods.
Approach
Synthesize a back view that inherits the reference image's texture via attention injection (K/V from front branch) under depth and text control, use it with a Zero-1-to-3 SDS coarse NeRF and DMTet fine mesh, and add a visibility-aware patch consistency loss to remove SDS color distortion in side regions.
Key results
- On THuman2.0, ConTex-Human reaches PSNR 21.4 and LPIPS 0.063 versus Magic123's 18.8 and 0.099, and PaMIR's 21.0 and 0.072.
- On SSHQ, ConTex-Human attains the best CLIP similarity of 0.903 for novel views, above PaMIR's 0.887 and Magic123's 0.882.
- Ablation on THuman2.0 shows removing back-view synthesis drops PSNR from 21.35 to 20.19 and raises LPIPS from 0.0634 to 0.0905 (CLIP 0.0778 vs 0.0634).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ConTex-Human instead of PIFu | single-image free-view human rendering · THuman2.0 | PSNR: 20.4 -> 21.4 | Helped |
| use ConTex-Human instead of PaMIR | single-image free-view human rendering · THuman2.0 | LPIPS: 0.072 -> 0.063 | Helped |
| use ConTex-Human instead of Magic123 | single-image free-view human rendering · THuman2.0 | CLIP: 0.910 -> 0.932 | Helped |
| use ConTex-Human instead of Magic123 | single-image free-view human rendering · SSHQ | CLIP: 0.882 -> 0.903 | Helped |
| add texture-consistent back-view synthesis (ablation) | ConTex-Human · THuman2.0 | PSNR: 20.188 -> 21.27 | Helped |
| add visibility-aware patch consistency loss (ablation) | ConTex-Human with back-view synthesis · THuman2.0 | LPIPS: 0.0658 -> 0.0634 | Helped |
Surprising findings
- Ablation metrics for the back-view and VPC components are inconsistent in places (e.g., 'with back & w/o VPC' has higher LPIPS than the full model), and SSIM for the proposed method is second-best, not best.
Limitations
- Coarse geometry in hand and foot regions; coarse-stage failures are not corrected by fine stage.
- Side and invisible regions are lower quality than front and back views and occasionally noisy.
- Person-specific optimization requiring over one hour of training per subject.
Experimental setup
Datasets: THuman2.0 SSHQ
Benchmarks: THuman2.0 SSHQ
Metrics: PSNR SSIM LPIPS CLIP similarity
Baselines: PIFu PaMIR Magic123
Builds on: DreamFusion Zero-1-to-3 Magic123 DMTet Instant-NGP ECON
Closest papers by meaning
- HumanRef: Single Image to 3D Human Generation via Reference-Guided Diffusion (CVPR 2024 · similarity 0.88)
- SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion (CVPR 2024 · similarity 0.86)
- GaussianIP: Identity-Preserving Realistic 3D Human Generation via Human-Centric Diffusion Prior (CVPR 2025 · similarity 0.86)
- PSHuman: Photorealistic Single-image 3D Human Reconstruction using Cross-Scale Multiview Diffusion and Explicit Remeshing (CVPR 2025 · similarity 0.86)
- SceneTex: High-Quality Texture Synthesis for Indoor Scenes via Diffusion Priors (CVPR 2024 · similarity 0.86)
- PaintHuman: Towards High-Fidelity Text-to-3D Human Texturing via Denoised Score Distillation (AAAI 2024 · similarity 0.86)
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.