Each frame will be broken into this number of substeps. Additional substeps are required for fast moving collisions or sudden forces.

The default substeps can be very aggressive, usually if the Vellum solver is too stretchy, raising substeps to 2 or 5 is a good first start.

Within each substep, this number of passes will be taken by the constraint enforcement operations. Stiff constraints can require more iterations to converge. A good starting point is the diameter of the geometry - the number of edges between the farthest points.

The default constraint iterations use a Gauss-Seidel approach that is fast to converge. However, if it doesn’t fully converge due to too high stiffness, or impossible configurations, it will leave the error as bad looking triangles. The smoothing iterations use a Jacobi approach which is slower to converge but leaves error spread out in a more attractive fashion. The default of ten passes helps smooth out error, but might need to be increased if the overall Constraint Iterations is very high.

Number of collision detection passes to perform. These are interleaved between the constraint iterations. Since collision is expensive, it is best to minimize this. But frequent interleaving helps avoid tent-poling effects where a small collider is fighting with the no-stretch constraints. In practice we find 10 to be correct for most situations, and substeps often being a better solution to increase quality.

After all constraints are performed, a final round of collision detection is done. Collisions are often the most noticeable failure mode, and it is ideal if the next frame can start with non-intersecting geometry. Thus a final cleanup pass can achieve these requirements. We have found that “number of stacked layers + 2” is a good estimate for this number. This allows the effect of the underlying collider to ripple through the stacked layers fully.

In any collision pass, any colliding pair may not be fully resolved. This number of additional collider-pair passes will be run until they are resolved. Since these are only performed on active colliders (and no new collision search is done) this is very cheap.

The layer integer point attribute is used to flag point as belonging to different layers of cloth. Higher numbers refer to higher layers. Layer Shock will make lower layers this many times heavier during collision evaluation, ensuring the higher layers will move out of their way. The rest of the dynamics are unaffected by this, and the difference is fixed regardless of the number of layers between the two. This can be thought of as a way to dial between one-way layering of sims and fully coupled sims.

A threshold at which to apply full friction. When the ratio of the tangential velocity and the normal impulse is less than this, the tangential velocity will be fully eliminated through friction. This is roughly tan() of the slope angle that will allow sliding under gravity.

If the static threshold fails, this controls what percentage the tangential velocity will be reduced in the dynamic friction case.

A scale factor on the amount of friction effect to apply for collisions with external geometry.

A scale factor on the amount of friction effect to apply for collisions with self geometry.

A scale factor for the volume collider’s static friction. Useful to create frictionless grounds.

A scale factor for the volume collider’s dynamic friction. Useful to create frictionless grounds.

If this checkbox is turned on, all constraints in the specified Constraint Group will be solved in a separate, interleaved pass to the remaining constraints.

There are two main uses for this option. First, there are some expensive constraint types that do not need to be solved as frequently as the rest of the constraints, so enabling this option and choosing a low Solve Frequency can improve performance. For example, cloth bend constraints on high resolution cloth with low bend stiffness such a silk or cotton. You need to ensure that you're solving distance constraints every pass so the cloth doesn’t stretch. However, bend constraints are expensive and don’t need to be as strong, since silk and cotton have very low bend resistance and wrinkle easily. Solving the bend constraints as a secondary pass will give you strong performance gains. However, this method isn’t practical for stiffer material such as leather.

The second use case is constraints that cause frequent changes in the constraint topology, which causes the solver to re-order the constraint solve, possibly causing jittering in the remaining constraints. An example is organic tissue represented by tetrahedra and attached together with sliding Stitch constraints. As the sliding causes changes in the constraint topology the tetrahedra might jitter slightly as the constraint solve order changes. Moving the Stitch constraints to another pass by setting Constraint Group to @type=ptprim and increasing the Solve Frequency to 1 will retain the same stiffness but eliminate tetrahedral jittering due to changes in the solve order.

The constraint group(s) that should be solved in the interleaved secondary pass. This parameter accepts standard group syntax, so can contain explicit group names and ad-hoc groups, often specifying one or more constraint types with something like @type==bend. Any constraint not contained in these groups will be solved in the primary pass as usual.

The frequency at which to solve the secondary pass. A setting of 1 will solve the secondary pass as often as the primary pass, while a setting of 0.25 will solve once every four passes of the primary pass (one quarter of the time).

The multi-pass options allow a substep to be repeated until certain conditions are met. The current conditions are designed to fix issues that are caused by disabled points causing geometry to be caught and cause the undisabled points to generate stretching. Since collisions always pre-empt constraints, the result is cloth or hair stretching. Points adjacent to auto-disabled points will be themselves disabled if too much stretching is detected. The solve step will then be repeated in the hopes this frees up the geometry.

The maximum number of times to repeat the substep. If no new points need to be disabled, the process will stop immediately.

The amount of stretch at the end of a solve step that will trigger points to be eligible for disabling. This is to detect when failed collisions are pulling the Vellum object apart. By failing additional points the object usually can be released and result in a better result than continuing to stretch.

If points fail to resolve their collisions after the post-collision passes, they will be flagged as disabled, allowing them and any primitives attached to them to move collision free.

If a disabled point detects it is no longer in a tangled situation, and is connected to non-disabled point, it will re-activate itself. The hope is that it has moved back to the correct side of the geometry.

The overlap_self and overlap_external attributes will be created and initialized to respect the initial setup of the solver.

Note if the attributes are already present, they will not be initialized. If the configuration has changed due to modeling operations you may need to delete those attributes.

The overlap_self and overlap_external will be updated to reflect the current configuration. The only decrease with this operation, so to initialize set to a large initial value.

The unshared faces of a tetrahedral mesh will generate collision geometry if this is set. Sometimes one wishes to use both a triangle shell and a tetrahedral mesh, making this operation redundant.

Various acceleration limiting options can be used to prevent the simulation from being over-eager to obey non-realistic forces. These can otherwise result in large energy spikes.

If Max Acceleration is exceeded for a point, assume it signals a sharp, discontinuous collision where the second order prediction will be wrong and add erroneous motion, often in the form of bouncing. In this case, fallback to first order integration for the affected point.

Cap the amount that velocity of a particle is allowed to change as a result of any of the dynamics. This is useful to prevent some instant motion being mis-identified as a massive force, and thus avoid fly-away particles.

During the collision resolution, if the collision correction moves a particle more than the acceleration amount, cap the effect. The hope is to fail more gracefully when a part of a model snags on the cloth and is pulling it in a surprising fashion.

When a weld breaks, the two new points start off next to each other. If the surfaces do not separate naturally due to whatever triggered the break, they may trigger collision detection and be pushed out, causing an explosive motion when breaking occurs. This option sets the disableself attribute when points are de-welded to avoid these self collisions. Note that this may cause layered cloth to self penetrate, however.

Normalize the stress computation over time, so that the computed values are more predictable as the Substeps parameter changes.

How frequently welds and constraints are tested to see if they have reached their breaking point. Higher rates will give more accuracy to the solve, but changes to the topology from a broken constraint can slow down the solve considerably.

The method used for finding the next closest position on the target geometry when sliding Attach to Geometry or Stitch Points constraints. Closest Point simply chooses the closest point on the target geometry to the projected sliding position. This approach is fast but can improperly jump across concavities in the target geometry. Traverse Polygons starts from the current target primitive and successively walks outwards, finding the closest point on the surrounding primitives. This approach is more expensive but handles concave target geometry better. Traverse Triangles (Optimized) is similar to the previous option in its improved handling of concavities, but can be many times faster as it uses specialized triangle distance functions. However, it can only be used with target geometry consisting of triangles.

Potential intersection particles are any within this scaled distance of the average of the two particles pscale attribute. This is an overestimate because usually collisions are not updated during the constraint iterations, so it needs to record not just the currently colliding particles, but those that may start to collide due to the earlier iterations.

This also effects the range of the attraction force in clumping.

The maximum number of particles that will be considered when searching for potential collisions over the substep. Capping this is useful to avoid excessive computations, if too many particles are created at one spot. This parameter is ignored if OpenCL Neighbor Search is enabled, in which case all neighbors are considered within the radius determined by Search Scale.

The pscale attribute is used to determine the radius of each particle. If all particles have the same radius, faster acceleration structures can be used to find neighbors.

Ignore any neighbors that have the same non-empty name or non-negative piece point attribute. This option can be enabled to create separate clusters of grains that only interact with other clusters, often in conjunction with a Shape Match constraint to give rigid behavior.

When multiple particles collide at the same time, this will average out all the constraints. This is effective in ensuring stability, but does not preserver momentum. Thus, when combined with internal forces, such as clumping, bunches of particles may accelerate under their own force.

A weighting for how much the particle collision forces are weighted. A value of zero will disable particle collision.

Scale with the repulsionweight point attribute.

How strongly particles are kept apart. Higher values result in less bouncy repulsion.

A weighting for how much the particles will naturally stick together when close. A value of zero will disable particle clumping.

Scale with the attractionweight point attribute.

How strongly nearby particles stick to each other. Higher values result in a less bouncy adhesion.

Artificially scales the mass of particles according to their position with respect to gravity. By making particles higher up lighter, stacks of particles will converge faster and be more stable.

The amount of scaling to perform. Higher numbers increase the contrast between successive particles. A value of 0 will cause no ratio between particles, a value of 1 a 15% ratio between two particles stacked vertically.

Too high a number makes higher particles extremely light and destabilizes the system.

The up-vector used to define a gradient of particle masses.

Should be in the direction of stacking, so usually is opposite to that of gravity.

The radius of the kernel used for fluid calculations. In general the default value of 4 is recommended, but it can be increased or reduced slightly for different effects. In particular this value can effect the size of the drops created by surface tension, with larger values creating larger drops and fewer individual particles. Note that larger values will result in slower simulation and more memory usage for neighbor lists. Also values much below 3 or above 5 can lead to instability.

The viscosity of the fluid. Low values for viscosity help keep the simulation stable, while higher values can simulate liquids like honey. A per-particle viscosity attribute can be used to multiply this value for variable viscosity. Fluid particles will different phase values will be solved independently for viscosity, allowing multi-phase fluid behavior.

The viscosity of the fluid in contact with an SDF collision. This value is only used when the solver has a ground plane or a collision object created with the VDB Collider SOP.

The type of viscosity solve to perform. Explicit is fast but can be unstable for high particle counts, high viscosity, or low substeps. Implicit is slower but remains stable even for higher particle counts and viscosity values.

The tolerance for the viscosity solve when using the Implicit solve. Lower values are more accurate at the cost of more iterations.

The maximum number of iterations for the Implicit viscosity solve.

The surface tension of the fluid. Higher values for this setting reduce the curvature of the fluid and cause it to form blobs. A per-particle surfacetension attribute can be used to multiply this value for variable surface tension. Fluid particles will different phase values will be solved independently for surface tension, allowing multi-phase fluid behavior.

The adhesion of the fluid. Higher values for this setting cause the fluid to stick to the normal surface of collisions, without any friction effects. You can increase Collision Viscosity to have collisions' velocity affect the fluid in a friction-like way. A per-particle adhesion attribute can be used to multiply this value for variable adhesion. Fluid particles will different phase values will be solved independently for adhesion, allowing multi-phase fluid behavior.

Vellum fluids and grains perform many calculations involving nearby points. Generally performance is greatly improved by ensuring that points that are near other in 3D space are also near each other in memory. This option enables a spatial sort of the particles at the specified interval of frames.

Specifies the type of solve to perform. Full provides the full Vellum feature set. The Minimal mode has reduced functionality, is optimized for fast Substeps, and is particularly useful for large grain and fluid simulations.

Specify any desired compile flags for the kernel.

When Finish Kernels is disabled, no attempt is to wait for the OpenCL kernels to complete before continuing the next solver. This lets them run in the background until their results are actually needed. To simplify debugging or timing, it is useful to ensure kernels are finished to make sure errors are detected in the right spot.

When loading kernels from disk the kernel is cached to avoid regenerating it every solve. Turning this on forces the re-loading and recompiling of the kernel. This is useful if #include files refer to code that has changed, or the kernel file is changed in an external text editor.

It should always be disabled when prototyping is complete.

When doing graph coloring, a fast parallel OpenCL algorithm is used. Unfortunately it may require 10× more memory than the rest of the solve on tetrahedral meshes. Thus, systems that might fit in memory for solving will not succeed the color pass. Disabling this forces all graph coloring to be done in a slower sequential manner, preserving RAM for the actual solve.

Perform any neighbor searches for grains and fluids using OpenCL, which is faster than the CPU but can use a bit more GPU memory.