How qiisun/meshflow works
MeshFlow is a PyTorch system for unconditional, equivariant flow-matching mesh generation, with category-conditioned demo controls. The main workflow is training mesh data into a learned model and checkpoint, then sampling that checkpoint into mesh files; users can also interactively generate and post-process meshes or evaluate generated outputs. The graph retains the indexed analysis, logging, and attention internals. Some subsystems are documented or indexed but not fully excerpted, so only documented or directly evidenced relationships are shown.
Main entry points: train.py Training loop · inference.py Mesh inference · equidit.py Equivariant DiT · gradio_demo.py Interactive demo
Components 19
Mesh data
- Mesh dataset
datasets/mesh_dataset.py - Mesh data files
- Noise matching
utils/ot_utils.py
Training and model
- Training loop
train.py - Flow matching
flow_matching.py - Equivariant DiT
models/equidit.py - Attention layers
models/attention.py - Model utilities
models/utils.py - Experiment logging
utils/logger.py
Generation
- Model checkpoint
- Mesh inference
inference.py - Generated meshes
- Interactive demo
gradio_demo.py - Mesh post-processing
gradio_demo.py
Evaluation and analysis
- Point evaluation
tools/point_evaluation.py - Mesh metrics
utils/mesh_evaluation_metrics.py - Sampling-step analysis
tools/plot_chamfer_vs_steps.py - Chamfer distance
utils/chamfer3D/dist_chamfer_3D.py
Other components
- User
Connections 24
- User → Training loop: starts training
README.md - Mesh data files → Mesh dataset: loaded for training
README.md - Mesh dataset → Noise matching: aligns noise
datasets/mesh_dataset.py - Training loop → Mesh dataset: loads batches
train.py - Training loop → Equivariant DiT: builds and trains
train.py - Training loop → Flow matching: uses flow matching
train.py - Training loop → Experiment logging: logs experiments
train.py - Equivariant DiT → Attention layers: uses layers
models/equidit.py - Equivariant DiT → Model utilities: uses utilities
models/equidit.py - Training loop → Model checkpoint: saves checkpoint
README.md - User → Mesh inference: requests generation
README.md - Model checkpoint → Mesh inference: loads weights
inference.py - Mesh inference → Equivariant DiT: samples with
inference.py - Mesh inference → Flow matching: integrates trajectory
inference.py - Mesh inference → Generated meshes: exports meshes
README.md - User → Interactive demo: selects generation
README.md - Interactive demo → Generated meshes: generates raw mesh
gradio_demo.py - Interactive demo → Mesh post-processing: runs post-processing
README.md - Mesh post-processing → Generated meshes: exports processed mesh
README.md - User → Point evaluation: evaluates outputs
README.md - Generated meshes → Point evaluation: provides generated meshes
README.md - Point evaluation → Mesh metrics: computes metrics
- User → Sampling-step analysis: analyzes sampling
tools/plot_chamfer_vs_steps.py - Sampling-step analysis → Chamfer distance: measures distance
tools/plot_chamfer_vs_steps.py