normalize joint weights #10539
normalize joint weights #10539
Conversation
robtfm
added
C-Feature
i can do this, but the warning becomes unavoidable ... if you load a gltf and once loaded you call up to you? |
|
Hmm 🤔 I think we can leave it. The real requirement here is for a general-purpose mesh-validation asset preprocessor IMO. |
| }; | ||
|
|
||
| for weights in joints.iter_mut() { | ||
| let sum: f32 = weights.iter().sum(); |
There was a problem hiding this comment.
Would it be worth verifying the weights are positive first? If someone accidentally uses with_inserted_attribute with a mixture of positive and negative floats, this will do weird things.
There was a problem hiding this comment.
negative weights are technically not allowed in the spec, whereas not summing to precisely 1 is allowed due to rounding issues.
but sure, no reason not to. i arbitrarily chose to clamp negatives to zero, i guess other treatments might be better (like if all are negative, maybe you should just flip the sign?) but since we're in invalid-data territory there's probably no perfect answer. if someone wants a different treatment they can always write their own normalize function.
# Objective allow automatic fixing of bad joint weights. fix bevyengine#10447 ## Solution - remove automatic normalization of vertexes with all zero joint weights. - add `Mesh::normalize_joint_weights` which fixes zero joint weights, and also ensures that all weights sum to 1. this is a manual call as it may be slow to apply to large skinned meshes, and is unnecessary if you have control over the source assets. note: this became a more significant problem with 0.12, as weights that are close to, but not exactly 1 now seem to use `Vec3::ZERO` for the unspecified weight, where previously they used the entity translation.
Objective
allow automatic fixing of bad joint weights.
fix #10447
Solution
Mesh::normalize_joint_weightswhich fixes zero joint weights, and also ensures that all weights sum to 1. this is a manual call as it may be slow to apply to large skinned meshes, and is unnecessary if you have control over the source assets.note: this became a more significant problem with 0.12, as weights that are close to, but not exactly 1 now seem to use
Vec3::ZEROfor the unspecified weight, where previously they used the entity translation.