Research map / Image generation / Autoregressive and latent generation
3DTopia-XL: Scaling High-quality 3D Asset Generation via Primitive Diffusion
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 3DTopia-XL, a native 3D diffusion model that generates PBR-textured 3D assets by running a DiT over a compact primitive-based representation (PrimX) compressed by a per-primitive patch VAE.
Problem
Generating high-quality 3D assets with geometric fidelity and physically based rendering (PBR) materials, with fast optimization, is hard; existing SDS, triplane reconstruction, and native 3D diffusion methods fall short on quality or PBR support.
Approach
Represent a textured mesh as N volumetric primitives anchored on the surface (SDF, albedo, metallic/roughness payloads in a N×D tensor), compress each primitive's voxel patch with a VAE, and model the set of latent primitives with a Transformer diffusion model (DiT) with cross-attention conditioning.
Key results
- PrimX fits a textured mesh in 1.5 minutes versus 10-16 minutes for baselines at a 1.05M parameter budget, with the lowest Chamfer Distance (1.310e-4 vs 4.502e-4 for the next best MLP).
- On text-to-3D with unseen prompts, 3DTopia-XL reaches CLIP Score 24.33 vs 22.54 for 3DTopia and 21.98 for Shap-E.
- Image-to-3D results are qualitative only (Figure 5), with the authors claiming best visual and geometry quality against LGM, InstantMesh, CRM, CraftsMan, Shap-E and LN3Diff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Represent the mesh with PrimX rather than the tested alternative representations | 3D representation fitting · 3D representation fitting | CD ×10−4: MLP 4.502; MLP w/ PE 4.638; Triplane 9.678; Dense Voxels 7.012; PrimX 1.310 | Helped |
| Represent the mesh with PrimX rather than the tested alternative representations | 3D representation fitting · 3D representation fitting | runtime: MLP 14 min; MLP w/ PE 14 min; Triplane 16 min; Dense Voxels 10 min; PrimX 1.5 min | Helped |
| Use 3DTopia-XL instead of ShapE or 3DTopia | Text-to-3D generation · Text-to-3D | CLIP Score: 21.98; 22.54; 24.33 | Helped |
| Use N=64 primitives with a=32 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F SDF: 62.52 -> 61.05 | Hurt |
| Use N=64 primitives with a=32 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F RGB: 24.23 -> 22.18 | Hurt |
| Use N=64 primitives with a=32 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F Mat: 18.53 -> 18.10 | Hurt |
| Use N=256 primitives with a=16 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F SDF: 62.52 -> 59.05 | Hurt |
| Use N=256 primitives with a=16 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F RGB: 24.23 -> 23.50 | Hurt |
| Use N=256 primitives with a=16 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F Mat: 18.53 -> 18.61 | Hurt |
| Use N=512 primitives with a=8 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F SDF: 62.52 -> 59.57 | Hurt |
| Use N=512 primitives with a=8 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F RGB: 24.23 -> 22.58 | Hurt |
| Use N=512 primitives with a=8 resolution instead of N=2048, a=8 (ablation) | PrimX · PrimX fitting | PSNR-F Mat: 18.53 -> 18.50 | No clear effect |
Surprising findings
- Triplane representation fits worse than a dense voxel and MLP baselines at the same parameter budget, with CD of 9.678e-4 versus 4.502e-4 for MLP.
Limitations
- Image-to-3D comparison is qualitative only, with no quantitative metrics reported in the main text.
- Text-to-3D evaluation relies only on CLIP Score, a coarse alignment metric.
- Training dataset is not described publicly, limiting reproducibility.
Experimental setup
Datasets: internal 3D training dataset of textured GLB meshes
Benchmarks: internal test set of GLB meshes (representation fitting) image-to-3D qualitative comparison text-to-3D unseen prompts
Metrics: Chamfer Distance PSNR (SDF, albedo, material) CLIP Score runtime
Baselines: LGM InstantMesh Real3D CRM CraftsMan Shap-E LN3Diff 3DTopia
Builds on: Diffusion Transformer Large Reconstruction Model M-SDF v-prediction Classifier-free guidance
Closest papers by meaning
- DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior (ICLR 2024 · similarity 0.89)
- One-2-3-45++: Fast Single Image to 3D Objects with Consistent Multi-View Generation and 3D Diffusion (CVPR 2024 · similarity 0.89)
- Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.88)
- Hi3DEval: Advancing 3D Generation Evaluation with Hierarchical Validity (NeurIPS 2025 · similarity 0.88)
- TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (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.