[Siggraph 2025] Intersection-free Garment Retargeting

This is the opensource reference implementation of the SIGGRAPH 2025 paper Intersection-free Garment Retargeting. This code is modified based on PolyFEM. It uses the nonlinear optimizers, linear solvers, and collision handling in PolyFEM.

Files

• src/: source code
• cmake/ and CMakeLists.txt: CMake files
• json-specs/: input JSON configuration specifications
• garment-data/: input data and scripts needed to run the code
• tests/: unit-tests

Build

The code can be compiled with

mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j4

gcc 7 or later is recommended.

Run

The folder garment-data contains four basic examples of retargeting garments to target avatars: foxgirl_skirt, Goblin_Jacket, Goblin_Jumpsuit, and Trex_Jacket. To run the examples, e.g. for foxgirl_skirt,

cd garment-data/foxgirl_skirt
../../build/PolyFEM_bin -j setup.json --max_threads 16 > log

Output Files

• step_avatar_<n>.obj: optimization sequence of the target avatar geometry
• step_garment_<n>.obj: optimization sequence of the retargeted garment surface
• sdf.obj: avatar surface converted from the signed distance field
• source_skeleton.obj: skeleton of the source garment
• target_skeleton.obj: skeleton of the target avatar
• target_avatar.obj: target avatar geometry
• projected_avatar.obj: projected target avatar on its skeleton

Script Settings

The specification of each JSON configuration is described in json-specs/input-spec.json.

Meshes

• avatar_mesh_path: the target avatar mesh
• garment_mesh_path: the source garment mesh
• source_skeleton_path: the skeleton edge mesh of the source garment in .obj format
• target_skeleton_path: the skeleton edge mesh of the target avatar in .obj format
• avatar_skin_weights_path: the skinning weights of the target avatar, a matrix of size #number_of_skeleton_nodes times #number_of_vertices (optional)

For avatars and garments, only .obj triangular mesh is supported. The source and target skeletons should share the same mesh connectivity, and skeleton joints should be ordered in the same way. The skinning weights will be used to project the target avatar to its skeleton. If avatar_skin_weights_path is not provided, the vertices will be simply projected to the closest bone in distance. Ideally, the source and target skeletons are in the same pose, so that the difference in pose does not cause unecessary geometric distortion on the garment.

Objectives and Weights

There are multiple objectives used in the method, each has its own weight:

• similarity_penalty_weight: $L_\text{surf}$ preserves the surface shape, always set to 1
• curvature_penalty_weight: $L_\text{curvature}$ preserves the curve curvature
• twist_penalty_weight: $L_\text{torsion}$ preserves the curve torsion
• curve_center_target_weight: $L_\text{pos}$ preserves the relative position of curve loops (e.g. hem, cuff)
• solver/contact/barrier_stiffness: $L_\text{contact}$ avoids contact by adding a barrier

The choice of barrier_stiffness depends on the mesh resolution, geometry scale, and the specific problem setup. The general rule is that when the min distance reported in the log is much smaller than dhat (say 1/100 of dhat), then it should probably be increased.

Optimizer

The nonlinear optimizer lives in a separate Github repository called PolySolve. The optimizer parameters are specified under JSON entry solver/nonlinear, whose specification is in nonlinear-solver-spec.json.