Research map / Image generation / Autoregressive and latent generation
MeshFlow: Efficient Artistic Mesh Generation via MeshVAE and Flow-based Diffusion Transformer
Weiyu Li, Antoine Toisoul, Tom Monnier, Roman Shapovalov, Rakesh Ranjan, Ping Tan, Andrea Vedaldi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TokenMerge queries instead of Q-Former learnable queries (ablation) | MeshVAE · MeshVAE ablation | F1 score: 49.47 -> 99.78 | Improved |
| use TokenMerge queries instead of FPS queries (ablation) | MeshVAE · MeshVAE ablation | F1 score: 60.18 -> 99.78 | Improved |
| use TokenMerge queries instead of Q-Former (ablation) | MeshVAE · MeshVAE ablation | Vertex distance (x100): 23.36 -> 0.75 | Improved |
| use TokenMerge queries instead of FPS (ablation) | MeshVAE · MeshVAE ablation | Vertex distance (x100): 18.29 -> 0.75 | Improved |
| increase downsampling ratio from 2 to 4 (ablation) | MeshVAE · MeshVAE ablation | F1 score: 92.65 -> 88.82 | Hurt |
| increase maximum vertex count from 2048 to 8192 (ablation) | MeshVAE · MeshVAE ablation | F1 score: 99.84 -> 99.71 | No clear effect |
| replace AR mesh generation with rectified flow on MeshVAE latents | MeshFlow vs FastMesh-V4K · Toys4K shape-conditioned generation | Chamfer distance (x100): 4.05 -> 2.33 | Improved |
| replace AR mesh generation with rectified flow on MeshVAE latents | MeshFlow vs FastMesh-V4K · Toys4K shape-conditioned generation | Hausdorff distance (x100): 10.22 -> 4.23 | Improved |
| use continuous MeshVAE latents instead of EdgeRunner ArAE tokens | MeshVAE · MeshVAE reconstruction | Compression ratio: 0.47 -> 0.014 | Improved |
| use continuous MeshVAE instead of EdgeRunner on reconstruction | MeshVAE · MeshVAE reconstruction | Chamfer distance (x100): 1.21 -> 1.29 | No clear effect |
| use rectified flow instead of FastMesh-V1K autoregression for single-mesh inference | MeshFlow · Single-mesh inference | Inference time (s): ~21 s -> ~1.2 s | Improved |
Closest papers by meaning
- EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation (ICLR 2025 · similarity 0.92)
- MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025 · similarity 0.90)
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.89)
- MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers (CVPR 2024 · similarity 0.89)
- TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing (CVPR 2025 · similarity 0.89)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.88)
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.
