Research map / 3D vision / 3D generation
StdGEN: Semantic-Decomposed 3D Character Generation from Single Images
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 StdGEN, a pipeline that generates semantically decomposed (body, clothes, hair) 3D anime characters from a single image in about three minutes, built around a feed-forward Semantic-aware Large Reconstruction Model (S-LRM).
Problem
Generating high-quality 3D characters from single images that are decomposable into semantic parts (body, clothing, hair), which prior holistic or slow SDS-based methods cannot provide efficiently.
Approach
Extends the LRM with a semantic field and a semantic-equivalent NeRF/SDF formulation so that multi-layer semantic surfaces can be extracted differentiably via FlexiCubes and trained with 2D supervision; adds an A-pose canonicalization and multi-view RGB/normal diffusion stage plus iterative multi-layer mesh refinement.
Key results
- On arbitrary-pose 3D generation, StdGEN reaches SSIM 0.916, LPIPS 0.084, FID 0.011 versus CharacterGen's 0.869/0.134/0.119 on the Anime3D++ test split.
- In a 28-person user study, StdGEN was preferred for overall quality 82.4% of the time versus 11.8% for the next best (Unique3D).
- The pipeline produces decomposed body, clothing and hair meshes in about three minutes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use StdGEN instead of CharacterGen for A-pose multi-view generation | Multi-view generation · A-pose multi-view generation | SSIM: 0.886 -> 0.958 | Helped |
| Use StdGEN instead of CharacterGen for A-pose multi-view generation | Multi-view generation · A-pose multi-view generation | LPIPS: 0.119 -> 0.038 | Helped |
| Use StdGEN instead of CharacterGen for arbitrary-pose multi-view generation | Multi-view generation · Arbitrary-pose multi-view generation | FID: 0.056 -> 0.014 | Helped |
| Use StdGEN instead of Unique3D for A-pose 3D character generation | 3D character generation · A-pose 3D character generation | SSIM: 0.889 -> 0.937 | Helped |
| Use StdGEN instead of CharacterGen for A-pose 3D character generation | 3D character generation · A-pose 3D character generation | LPIPS: 0.124 -> 0.066 | Helped |
| Use StdGEN instead of Unique3D for arbitrary-pose 3D character generation | 3D character generation · Arbitrary-pose 3D character generation | FID: 0.042 -> 0.011 | Helped |
| Use StdGEN instead of Unique3D for arbitrary-pose 3D character generation | 3D character generation · Arbitrary-pose 3D character generation | CLIP similarity: 0.903 -> 0.936 | Helped |
| Select StdGEN as best rather than CharacterGen | 3D character generation · User study | Overall quality selections: 4.0% -> 82.4% | Helped |
Limitations
- Evaluated mainly on anime characters from a single curated dataset
- Decomposition limited to base body, clothing and hair layers
- Relies on diffusion-generated multi-view images and normals, so errors there propagate
Experimental setup
Datasets: Anime3D++ Anime3D
Benchmarks: Anime3D++ test split
Metrics: SSIM LPIPS FID CLIP similarity
Baselines: CharacterGen InstantMesh Unique3D LGM OpenLRM Magic123 ImageDream SyncDreamer Zero-1-to-3 Era3D
Builds on: InstantMesh Era3D CharacterGen Zero-1-to-3 FlexiCubes
Closest papers by meaning
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.87)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.86)
- 3DToonify: Creating Your High-Fidelity 3D Stylized Avatar Easily from 2D Portrait Images (CVPR 2024 · similarity 0.86)
- Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.86)
- Progressive3D: Progressively Local Editing for Text-to-3D Content Creation with Complex Semantic Prompts (ICLR 2024 · similarity 0.86)
- DRiVE: Diffusion-based Rigging Empowers Generation of Versatile and Expressive Characters (CVPR 2025 · 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.