How yizhou42/masiv works
MASIV identifies geometry and material physics from dynamic video: users prepare a sequence, train a dynamic Gaussian reconstruction and physical estimator, then predict novel states or inspect trajectories. The graph preserves the documented generalization and visualization branches as well as the MPM/material and differentiable-rendering layers. Some script bodies and subsystems were not excerpted; their documented capabilities are shown, but uncertain internal links are omitted.
Main entry points: estimator.py Physical estimator · train_dynamic.py Dynamic training · __init__.py Scene loading · prepare_pacnerf_data.py Dataset preparation
Components 22
Video and scene data
- Video sequences
- Dataset preparation
prepare_pacnerf_data.py - Foreground matting
matting/model.py - Scene loading
scene/__init__.py - Dataset readers
scene/dataset_readers.py - COLMAP parsing
scene/colmap_loader.py
Gaussian reconstruction
- Gaussian model
scene/gaussian_model.py - Gaussian renderer
gaussian_renderer/__init__.py - Gaussian optimization
train_gs.py
Physics identification
- Dynamic training
train_dynamic.py - Physical estimator
simulator/estimator.py - MPM simulation
nclaw/sim/mpm.py - Material models
nclaw/material/meta.py - Constitutive laws
simulator/constitution.py
Prediction and inspection
- Future prediction
predict.py - Novel-condition tests
generalize.py - Trajectory visualization
visualize.py
Numerical components
- Differentiable SVD
nclaw/warp/svd.py - Warp autodiff tape
nclaw/warp/tape.py - Training losses
utils/loss_utils.py - Gaussian regularization
utils/reg_utils.py
Other components
- Researcher
Connections 23
- Researcher → Video sequences: provides
- Researcher → Dataset preparation: runs preprocessing
- Dataset preparation → Foreground matting
- Researcher → Dynamic training: starts training
- Video sequences → Scene loading: supplies sequence
- Scene loading → Dataset readers: dispatches loading
scene/__init__.py - Dataset readers → COLMAP parsing: reads COLMAP
scene/dataset_readers.py - Scene loading → Gaussian model: initializes model
scene/__init__.py - Dynamic training → Scene loading: loads scene
train_dynamic.py - Dynamic training → Physical estimator: optimizes physics
train_dynamic.py - Dynamic training → Gaussian optimization: uses optimization
train_dynamic.py - Gaussian renderer → Gaussian model: reads Gaussian state
gaussian_renderer/__init__.py - Physical estimator → Gaussian renderer: renders observations
simulator/estimator.py - Physical estimator → MPM simulation: simulates particles
simulator/estimator.py - Physical estimator → Material models: uses material model
simulator/estimator.py - Physical estimator → Constitutive laws: uses constitutive law
simulator/estimator.py - Material models → Differentiable SVD: uses decomposition
- Differentiable SVD → Warp autodiff tape: records gradients
nclaw/warp/svd.py - Physical estimator → Training losses: computes objectives
simulator/estimator.py - Researcher → Future prediction: runs prediction
- Researcher → Novel-condition tests: specifies novel conditions
- Researcher → Trajectory visualization: selects trajectory view
- Trajectory visualization → Physical estimator: loads simulation model
visualize.py