Single-image 3D reconstruction with cached flow-matching sampler

SAM 3D Objects reconstructs full 3D shape, texture, and layout from a single masked image. This fork adds HiCache to the slat-stage flow-matching sampler, skipping most backbone evaluations by forecasting the cached velocity with a scaled-Hermite polynomial.

mamba env create -f environments/default.yml
mamba activate sam3d-objects

# for pytorch/cuda dependencies
export PIP_EXTRA_INDEX_URL="https://pypi.ngc.nvidia.com https://download.pytorch.org/whl/cu121"

# install sam3d-objects and core dependencies
pip install -e '.[dev]'
pip install -e '.[p3d]' # pytorch3d dependency on pytorch is broken, this 2-step approach solves it

# for inference
export PIP_FIND_LINKS="https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.5.1_cu121.html"
pip install -e '.[inference]'

# CPU contract tests (no GPU or model weights)
pytest tests/ # expect 4 passed

# accel-module CPU unit self-test from the README
python -m sam3d_objects.model.backbone.generator.flow_matching.accel