blend shapes vs bones for facial rigs — when does each actually make sense and why is it never simple

The skeleton lock point matters and I'd push it further: bind pose changes, not adding bones, just adjusting rest rotations on existing ones, can quietly invalidate correctives that were sculpted against the old deformation. We had a case where a rigger improved shoulder range by tweaking rest orientation on the clavicle and upper arm. Technically the same skeleton, but the deltoid correctives we'd built were now fighting the new deformation envelope. Took a while to even diagnose because the correctives still blended in. They just looked subtly wrong in ways that were hard to articulate until we compared against reference footage directly.

So the lock really needs to cover skeleton and bind pose, not just topology and hierarchy. Both of those define the deformation space the correctives live in.

The topology invalidating correctives is exactly why we don't let anyone sculpt correctives until the mesh has passed final polycount review and concept is signed off. It adds a gate but it's saved us from the "we moved the eye socket 2mm and now 11 correctives look subtly wrong" situation more than once. The hard part is convincing people to wait, because everyone wants to start the "fun" sculpting work before the boring approval stuff is done.

Also if you're in Blender: the corrective smooth modifier driven by shape key influence can fake some correctives procedurally, which at least cuts down on the number of hand-sculpted ones you're maintaining.

The version-lock is a good call, though we found we had to extend it to the skeleton hierarchy as well. Base mesh lock without a matching skeleton lock still left us exposed if someone added a bone for gameplay reasons or adjusted joint placement late. A single corrective can reference both mesh and bone positions implicitly, so either change will break it.

We ended up coupling both locks under a shared version token in our asset pipeline. A bit annoying to set up, but the alternative is discovering broken correctives during polish when nobody wants to touch the rig. Much worse time to find out.

The polycount gate is a solid process. We do something similar but also version-lock the base mesh in the asset repo once corrective authoring begins. No topology changes without a formal sign-off that includes whoever owns the rig. What's killed us in the past isn't one big mesh change, it's creeping small edits: someone pulls a vert to fix a silhouette, nobody considers it a topology change, and suddenly two correctives are broken in ways that don't show up until a camera close-up in review. The gate has to cover those micro-changes too or it doesn't really hold.

oh this is painfully familiar. we had a shoulder corrective that was fine for months and then someone did a "cleanup pass" on rest rotations, like 4 degrees on the clavicle, totally reasonable change, and suddenly a completely unrelated cheek puff shape started pulling sideways. took us two days to figure out what happened because nothing in the commit message flagged it as a rig change.

we ended up putting corrective authoring in a separate scene file that references the skeleton as a linked asset, so a rest pose change can't even save without breaking the reference. it adds friction to legitimate skeleton edits but at least the damage is immediately obvious rather than showing up three weeks later in a render.

The corrective shape maintenance cost is real and it compounds. Every meaningful change to the base mesh can invalidate correctives that were sculpted against the old topology, and there's no automatic way to know which ones broke. You find out when you see a weird pinch in a pose you thought was fixed. Bones don't have this problem the same way: adjust the rig, the skin weights accommodate, and corrective-equivalent behavior comes from rig logic rather than stored deltas. Blend shapes pay off when the character is locked and stable. If the mesh is still iterating, bones are a lot more forgiving over time.

The bind pose point is real and it's especially sneaky because nothing warns you. A corrective sculpted against a specific rest pose looks totally fine until someone adjusts a wrist rotation by two degrees to fix an unrelated clipping issue, and suddenly your lip corner corrective has a subtle artifact that only shows up at certain camera angles. It took me an embarrassingly long time to track down the first time it happened. I kept assuming the problem was the corrective itself, not the skeleton it was sculpted against.

The practical fix we landed on: treat bind pose as part of the rig spec, not just the mesh spec. Any rest rotation change goes through the same sign-off process as a topology change. Sounds bureaucratic but it's saved us more than once.

oh this is so real. had a wrist corrective that looked fine for six months and then an animator "cleaned up" some rest rotations, like 3 degrees on the shoulder, and suddenly a mouth corner was pulling wrong on extreme expressions. took an embarrassingly long time to connect those two things because the corrective was on the face and nobody was thinking about the shoulder.

these cross-rig dependencies are genuinely invisible until they break. the bind pose lock is the right call but it's also the kind of process gate that's really hard to sell to a team that just wants to iterate fast.