mocap retargeting onto stylized rigs: when does the data just stop being useful

This is where I landed after fighting the retargeter for too long. What finally clicked was realizing mocap carries timing and weight information far better than it carries literal joint angles for stylized proportions. Once I started treating it as motion reference rather than motion data, blocking in broad timing from a reference layer and animating on top, quality improved immediately. You also get faster polish passes because you're not untangling retargeting artifacts on top of everything else. On stylized characters especially, that cleanup time adds up.

Completely agree on the reference skeleton quality point. I was using a HIK biped I'd set up years ago and never thought to revisit it, and my retargeting results were consistently mediocre in ways I kept blaming on the mocap data itself. Rebuilt the reference skeleton from scratch using a properly proportioned capture reference, and the difference was immediate. Noticeably less foot sliding, cleaner shoulder rotation, spine movement that actually read as intentional rather than jittery.

It's the kind of thing you assume the software handles transparently, but it really doesn't. Garbage in, garbage out, just one step earlier in the pipeline than you'd expect.

I'd push back on the proportions-get-overstated framing. The issue isn't purely visual. The rotation values themselves can become anatomically implausible when applied to a skeleton with significantly different limb ratios. Elbows hyperextending, shoulders riding up into the neck, hips rotating past any reasonable range. That's not a visual oddity, that's contact errors and IK failures that need correction passes regardless of what you were trying to extract from the data.

The rotations are kind of correct, but they're correct relative to the capture skeleton's geometry and constraints. On a stylized rig where the same joints have different physical limits baked in, that breaks down fast. Past a certain proportion delta you're really just keeping the timing data and reconstructing the poses anyway.

The reach settings in MotionBuilder's HIK solver are worth tuning alongside bone length compensation. I had a very similar issue with elongated leg proportions and dropping the full-body reach for the legs by around 20% made a noticeable difference. The default is calibrated for realistic skeletons and tends to over-extend on stylized rigs.

Also worth auditing your reference T-pose. If the proportions in your MB reference pose don't match your game rig reasonably closely, the compensation math is working from a bad baseline before you even touch reach settings. Spent longer than I'd like to admit figuring that one out.

Completely agree. I spent the better part of a week trying to get a Rokoko take onto a character with roughly 1.5x normal leg length. MotionBuilder's retargeter has bone length compensation settings, but even maxed out you get this floating hover on flat ground because the stride length in the capture was calibrated for a much shorter hip height. Ended up manually scaling root motion and redoing foot contact corrections by hand. At that point you've spent more time cleaning the retargeted data than you would have keyframing the broad strokes and using mocap only for fine detail on a properly proportioned rig.

It compounds too. It's not just the legs either. Spine and shoulder timing reads wrong because the pelvis offset from the ground is incorrect relative to everything else hanging off it.

I'd push back on this a bit. Proportions matter more than the "it depends on what you need" framing suggests. Rotation data from mocap encodes joint angles that make physical sense on the skeleton they were captured on. When you retarget to a character with a torso that's 40% longer, those same rotations don't produce equivalent body language, not because the retargeting is broken, but because the visual weight and leverage are completely different. A subtle spine lean that reads as casual confidence on a realistic biped can look like a full-body crunch on an elongated character.

The question isn't just whether you need naturalism. Even stylized animation benefits from mocap for timing and weight reference, but extreme proportion mismatches can scramble both of those things. I've had takes that looked perfectly fine in isolation but felt completely wrong on the actual character in context, and the fix was always dealing with the proportions, not cleaning the mocap data further.

Rebuilding from scratch is the right call, and the thing that most consistently kills HIK retargeting quality is non-zero rotations baked into the bind pose of the reference skeleton. If any bone in your T-pose reference has rotation values in the transform, the solver has to compensate on every frame, and it compounds badly in the spine and shoulders where small angular errors stack up. Every bone in the reference should be rotation-zero in T-pose, with orientation encoded purely in the joint orientation attribute, not the transform rotation.

Had an old HIK biped with accumulated 2–3 degree offsets scattered through the mid-spine joints that I'd never noticed. Zeroing everything and resetting the character definition improved results immediately in ways I hadn't been able to get out of the solver settings alone.