Rigged LOD: decimated clothing sinks into the body (bodies poke through clothes under VRAM pressure) #611

New issue

Open

opened 2026-10-10 01:15:49 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-10 01:15:49 +00:00

Owner

Causes

Simplification shrinks surfaces inward. Removing vertices replaces curved surfaces with chords inside the original. A decimated body shrinks under its clothes (harmless); a decimated garment shrinks into the body. Clothing fitted a few mm over skin breaks at a few mm of error.
Per-mesh decisions. Body, shirt and jacket each pick their own decimated step and their own creator-LOD cap, so clothing can drop a level before the body; clothing creators' low LODs are usually far worse than body creators'.

Fix

Avatar-coherent LOD: one creator-LOD level and one decimation level per avatar, for all of its rigged meshes (LL already uses one creator LOD per avatar for rigged).
Outward-safe steps for outer layers: offset each simplified step outward along its normals by its measured error so a reduced garment never sinks below its original surface; the innermost body layer keeps plain (inward) simplification. Needs a body vs clothing classification per avatar (e.g. the largest mesh / the one sharing the base skeleton with alpha-mask faces); cheap fallback: inflate everything but the largest mesh.
Keep the error-driven step choice, cache, shadow LOD and RT replay unchanged.

Related: #559, #561, #573, the un-subdivide ticket, the trigger rework ticket.

Under rigged-memory pressure (8 GB cards in crowds: skinned-vertex capacity squeezed by the VRAM ladder, decimated steps + creator-LOD caps engaged), worn clothing visibly drops in quality before the body, and bodies poke through clothes (reported on an RTX 3070 / 7.2 GB budget at a club; screenshot from the owner). Avatars that alpha their bodies under clothing hide it; many don't. ## Causes 1. **Simplification shrinks surfaces inward.** Removing vertices replaces curved surfaces with chords inside the original. A decimated body shrinks under its clothes (harmless); a decimated garment shrinks into the body. Clothing fitted a few mm over skin breaks at a few mm of error. 2. **Per-mesh decisions.** Body, shirt and jacket each pick their own decimated step and their own creator-LOD cap, so clothing can drop a level before the body; clothing creators' low LODs are usually far worse than body creators'. ## Fix - **Avatar-coherent LOD:** one creator-LOD level and one decimation level per avatar, for all of its rigged meshes (LL already uses one creator LOD per avatar for rigged). - **Outward-safe steps for outer layers:** offset each simplified step outward along its normals by its measured error so a reduced garment never sinks below its original surface; the innermost body layer keeps plain (inward) simplification. Needs a body vs clothing classification per avatar (e.g. the largest mesh / the one sharing the base skeleton with alpha-mask faces); cheap fallback: inflate everything but the largest mesh. - Keep the error-driven step choice, cache, shadow LOD and RT replay unchanged. Related: #559, #561, #573, the un-subdivide ticket, the trigger rework ticket.

sky_line added the

labels

2026-10-10 01:15:49 +00:00

sky_line added this to the Slipstream project

2026-10-10 01:15:49 +00:00

sky_line referenced this issue