USD Shape IO Support #101285

New Issue

Michael Kowalski · 2022-09-23T02:57:47+02:00

Michael Kowalski commented

2022-09-23 02:57:47 +02:00

On the most basic level, this feature would allow importing USD shapes (capsule, cone, cube, cylinder, sphere) as Blender meshes. More advanced implementations could include options for alternate, lightweight representations of shapes in the Blender scene, e.g., exploiting instancing, custom viewport rendering that does not require importing the mesh data, etc.

@charlesfleche discussed the following suggestions re USD shape support on the pipeline-assets-io-module Blender chat channel on 22 Sep 2022.

Our main use case is to represent collision objects. For performance reasons we can feed simple collision objects to the physic component of the game engine. For example, rather that computing a collision with a complex character head that as many faces, you can define a simple sphere that is much faster to compute collision against.

But sometimes simple shapes are not enough: in this case we still want as fast as a collision computation as possible, and use a convex mesh to approximate the visual shape as much as possible. And when we really, really require high precision we'd use a mesh that can be complex and convex.

We also have the concept of collision Level Of Detail. The typical example is a fence:

you don't the player to walk through the fence, but for that purpose you can approximate the shape of the fence with a single parametric box
but you still want a bullet to go through the holes of the fence: for collision with ammunitions, you'd use a more detailed mesh

But not all DCC support parametric shapes, so for simplicity (and for now), we are still importing USD shapes as meshes in the DCCs, but on way back to USD we compute the bounding volumes of the requested shape. In practice, it looks like this:

In USD: UsdGeomCube MyObject_BOX with a size attribute + xform
In the DCC: we extract the dimension of the box from the scale component of the transform and generate a box mesh, with custom code ran at import time
DCC user is editing the box
In the DCC: the object is a _BOX, we compute the Box Bounding Volume and store the x, y, z dimensions as in the xform as scale
In USD: UsdGeomCube MyObject_BOX wih a size attribute + xform

It is not great because the users don't have the same representation in engine vs the DCC. In engine they'd handle Radius and Height parameters, in the DCC it's a bunch of vertices.

So the idea, for Blender, would be to leverage the geometry nodes to give the user parameters to play with (radius, height), even if the actual object in Blender is still a mesh.

Geometry Nodes presets should be enough for now. After all, you can already generate those shapes in the Geometry Nodes already. What is important would be to have control, even if purely manual at import time, on how the Shape will be presented in Blender. For example, most users will want a Box to be defined with three parameters: sizeX, sizeY, sizeZ. However, depending of the physics engine you are dealing with, sometimes boxes are represented by two vectors: bottom-left and top-right. To match the Engine representation in the DCC, we'd need to be able to replace the default Geometry node by ours, where we'd expose the bottom-left / top-right vectors to the user.

On the most basic level, this feature would allow importing USD shapes (capsule, cone, cube, cylinder, sphere) as Blender meshes. More advanced implementations could include options for alternate, lightweight representations of shapes in the Blender scene, e.g., exploiting instancing, custom viewport rendering that does not require importing the mesh data, etc. @charlesfleche discussed the following suggestions re USD shape support on the `pipeline-assets-io-module` Blender chat channel on 22 Sep 2022. --- Our main use case is to represent collision objects. For performance reasons we can feed simple collision objects to the physic component of the game engine. For example, rather that computing a collision with a complex character head that as many faces, you can define a simple sphere that is much faster to compute collision against. But sometimes simple shapes are not enough: in this case we still want as fast as a collision computation as possible, and use a convex mesh to approximate the visual shape as much as possible. And when we really, really require high precision we'd use a mesh that can be complex and convex. We also have the concept of collision Level Of Detail. The typical example is a fence: - you don't the player to walk through the fence, but for that purpose you can approximate the shape of the fence with a single parametric box - but you still want a bullet to go through the holes of the fence: for collision with ammunitions, you'd use a more detailed mesh But not all DCC support parametric shapes, so for simplicity (and for now), we are still importing USD shapes as meshes in the DCCs, but on way back to USD we compute the bounding volumes of the requested shape. In practice, it looks like this: - In USD: `UsdGeomCube` MyObject_BOX with a size attribute + xform - In the DCC: we extract the dimension of the box from the scale component of the transform and generate a box mesh, with custom code ran at import time - DCC user is editing the box - In the DCC: the object is a _BOX, we compute the Box Bounding Volume and store the x, y, z dimensions as in the xform as scale - In USD: `UsdGeomCube` MyObject_BOX wih a size attribute + xform It is not great because the users don't have the same representation in engine vs the DCC. In engine they'd handle Radius and Height parameters, in the DCC it's a bunch of vertices. So the idea, for Blender, would be to leverage the geometry nodes to give the user parameters to play with (radius, height), even if the actual object in Blender is still a mesh. `Geometry Nodes` presets should be enough for now. After all, you can already generate those shapes in the `Geometry Nodes` already. What is important would be to have control, even if purely manual at import time, on how the Shape will be presented in Blender. For example, most users will want a Box to be defined with three parameters: `sizeX, sizeY, sizeZ`. However, depending of the physics engine you are dealing with, sometimes boxes are represented by two vectors: `bottom-left` and `top-right`. To match the Engine representation in the DCC, we'd need to be able to replace the default Geometry node by ours, where we'd expose the `bottom-left` / `top-right` vectors to the user.

Michael Kowalski commented

2022-09-23 02:57:47 +02:00

Added subscribers: @charlesfleche, @makowalski, @CharlesWardlaw, @SonnyCampbell_Unity, @dr.sybren

Michael Kowalski commented

2022-09-23 03:04:23 +02:00

Added subscriber: @BrianSavery

Zhen Dai commented

2022-09-23 07:36:05 +02:00

Added subscriber: @Zhen-Dai

Philipp Oeser commented

2022-09-23 10:30:01 +02:00

Added subscriber: @lichtwerk

Paul Larson commented

2022-09-28 08:03:50 +02:00

Added subscriber: @GeorgiaPacific

Michael Kowalski commented

2022-10-06 18:28:16 +02:00

Added subscriber: @mont29

Charles Wardlaw commented

2022-10-06 19:41:24 +02:00

Hey all,

I've made a patch for this internally. Currently it follows the Maya implementation, converting the UsdShape objects into mesh objects using the Adapter classes provided. At the moment it's one resolution fits all, but I could add a subdivision modifier or try to match subdivision schema on the file being imported. There is currently no round-tripping.

The Geometry Nodes idea is a good one with regards to attempting to round-trip shape changes, and that's something I could do, but how do folks want to detect that a geometry nodes network is mimicking Geometry Nodes vs not? Would export be based on network name + available attributes, and if both don't exist it gets treated like a regular mesh?

I suppose this would also require a new menu be added to the Shift+A menu for UsdShapes?

Hey all, I've made a patch for this internally. Currently it follows the Maya implementation, converting the UsdShape objects into mesh objects using the Adapter classes provided. At the moment it's one resolution fits all, but I could add a subdivision modifier or try to match subdivision schema on the file being imported. There is currently no round-tripping. The Geometry Nodes idea is a good one with regards to attempting to round-trip shape changes, and that's something I could do, but how do folks want to detect that a geometry nodes network is mimicking Geometry Nodes vs not? Would export be based on network name + available attributes, and if both don't exist it gets treated like a regular mesh? I suppose this would also require a new menu be added to the Shift+A menu for UsdShapes?

Bastien Montagne added this to the Pipeline, Assets & IO project 2023-02-09 15:39:30 +01:00

Bastien Montagne modified the project from Pipeline, Assets & IO to USD

2023-02-10 11:08:11 +01:00

Bastien Montagne added

and removed

labels 2023-02-10 12:24:54 +01:00

Sign in to join this conversation.