Speed up bone merge in export prep #75

Merged

kkbp-dev merged 1 commit from sialim/KKBP_Importer:master into development

2026-06-26 19:08:24 +00:00

sialim commented

2026-06-20 07:34:03 +00:00

Contributor

The "Prep for target application" merge gets extremely slow on high-bone-count models. On a 89-object model with ~600–1000 bones each it runs for hours, sometimes effectively days.

Cause: merge_weights calls mix_weights once per (mesh, bone) pair, and each call creates a VERTEX_WEIGHT_MIX modifier and runs bpy.ops.object.modifier_apply. Every modifier_apply forces a full depsgraph re-evaluation and an undo push, so the work scales with meshes × bones × per-apply overhead and is entirely main-thread bound (extra CPU cores don't help).

Fix: Do the same operation directly on vertex-group data. A VERTEX_WEIGHT_MIX in ADD mode is just target_weight += source_weight for vertices in the source group, followed by deleting the source group — so this rewrites the merge to accumulate weights with vertex_group.add(..., 'ADD') in a single pass per mesh. No modifiers, no operators, no per-bone depsgraph evaluation. Cost is now roughly proportional to total vertex count.
I verified the direct-add output is numerically identical to the modifier-based result.

Tested on this card (https://db.bepis.moe/koikatsu/view/392842), VRM + "very simple" preset: merge dropped from ~[3-4 hours] to ~[30 seconds]. Output weights unchanged.

The "Prep for target application" merge gets extremely slow on high-bone-count models. On a 89-object model with ~600–1000 bones each it runs for hours, sometimes effectively days. Cause: merge_weights calls mix_weights once per (mesh, bone) pair, and each call creates a VERTEX_WEIGHT_MIX modifier and runs bpy.ops.object.modifier_apply. Every modifier_apply forces a full depsgraph re-evaluation and an undo push, so the work scales with meshes × bones × per-apply overhead and is entirely main-thread bound (extra CPU cores don't help). Fix: Do the same operation directly on vertex-group data. A VERTEX_WEIGHT_MIX in ADD mode is just target_weight += source_weight for vertices in the source group, followed by deleting the source group — so this rewrites the merge to accumulate weights with vertex_group.add(..., 'ADD') in a single pass per mesh. No modifiers, no operators, no per-bone depsgraph evaluation. Cost is now roughly proportional to total vertex count. I verified the direct-add output is numerically identical to the modifier-based result. Tested on this card (https://db.bepis.moe/koikatsu/view/392842), VRM + "very simple" preset: merge dropped from ~[3-4 hours] to ~[30 seconds]. Output weights unchanged.

sialim added 1 commit

2026-06-20 07:34:04 +00:00

Speed up bone merge in export prep 3975f65ee0

The "Prep for target application" merge gets extremely slow on high-bone-count models. On a 89-object model with ~600–1000 bones each it runs for hours, sometimes effectively days.

Cause: merge_weights calls mix_weights once per (mesh, bone) pair, and each call creates a VERTEX_WEIGHT_MIX modifier and runs bpy.ops.object.modifier_apply. Every modifier_apply forces a full depsgraph re-evaluation and an undo push, so the work scales with meshes × bones × per-apply overhead and is entirely main-thread bound (extra CPU cores don't help).

Fix: Do the same operation directly on vertex-group data. A VERTEX_WEIGHT_MIX in ADD mode is just target_weight += source_weight for vertices in the source group, followed by deleting the source group — so this rewrites the merge to accumulate weights with vertex_group.add(..., 'ADD') in a single pass per mesh. No modifiers, no operators, no per-bone depsgraph evaluation. Cost is now roughly proportional to total vertex count.
I verified the direct-add output is numerically identical to the modifier-based result.

Tested on this card (https://db.bepis.moe/koikatsu/view/392842), VRM + "very simple" preset: merge dropped from ~[3-4 hours] to ~[30 seconds]. Output weights unchanged.

kkbp-dev commented

2026-06-26 19:07:28 +00:00

Owner

This is smart! Waiting even two minutes for a few bones to merge was a pain, and like you said the runtime goes from a coffee break to an unbearable wait when the amount of objects and bones increase. Thanks for diagnosing and finding an improvement!

❤️ 1

kkbp-dev changed target branch from master to development

2026-06-26 19:07:47 +00:00

kkbp-dev merged commit 4a0908d541 into development

2026-06-26 19:08:24 +00:00

kkbp-dev referenced this pull request from a commit

2026-06-26 19:08:24 +00:00

Merge pull request 'Speed up bone merge in export prep' (#75) from sialim/KKBP_Importer:master into development