You ran your video through an AI mocap tool, dropped it on your character, and it looks wrong — feet skating across the floor, a low-grade shiver in every joint, ankles sinking through the ground, knees snapping the wrong way. None of that means you did anything incorrectly. It is the normal output of a markerless solve, and it is fixable. This guide explains why these artifacts happen and walks through the exact cleanup workflow we use to turn raw AI capture into animation that holds up in a game or on screen.
Why AI mocap produces these errors
Markerless capture estimates a 3D skeleton from 2D pixels, and that translation is where the problems are born:
- 2D-to-3D solve: the AI infers a full 3D pose from a flat image. Tiny frame-to-frame disagreements in that inference show up as jitter.
- Depth ambiguity: a single camera cannot truly see how far away a limb is, so the solver guesses depth. Wrong guesses become foot-slide and ground penetration.
- Occlusion: when an arm crosses the torso or one leg hides the other, the solver loses the joint and invents a position — that is where limb and knee pops come from.
Multi-camera capture reduces all of this because triangulation resolves depth, but even multi-cam solves need cleanup. The artifacts are inherent to the method, not a sign of bad input.
Fixing foot slide
Foot-slide is the most obvious tell. The fix is contact-based, not cosmetic:
- Identify the frames where each foot is planted by inspecting toe and heel velocity.
- Lock those contacts so the foot stays fixed in world space during the plant.
- Use IK to keep the leg chain attached to the locked foot while the hips move naturally over it.
Done right, the character now pushes off the ground instead of gliding over it — the single biggest upgrade in believability.
Killing jitter without killing the motion
Jitter is high-frequency noise on top of real movement. The trap is over-smoothing: crank a filter too hard and you flatten the snap and weight out of the performance. The approach:
- Apply gentle smoothing or a low-pass filter targeted at the noisy frequency band, not the whole curve.
- Protect keyframe extremes — the contact, the impact, the held pose — so the motion keeps its punch.
- Smooth per-joint, not globally; fingers and head usually need more help than hips.
Ground penetration and joint pops
Ground penetration — ankles or toes dipping below the floor — is fixed by establishing a true ground plane and clamping foot height to it, then re-checking the IK so the legs do not over-straighten. Limb pops at the knees and elbows usually come from the solver flipping a joint past its natural range or guessing through occlusion. The repair is to:
- Find the frames where the joint angle snaps or hyperextends.
- Constrain the joint to a sane range and correct the rotation axis so the bend points the right way.
- Blend the corrected frames into the surrounding motion so the fix is invisible.
The cleanup workflow, step by step
- Assess: review the raw solve and catalogue the artifacts and which frames are worst.
- Ground and contacts: set the floor plane, lock foot contacts, fix penetration.
- Filter: remove jitter per-joint while protecting the key poses.
- Joint repair: fix knee and elbow pops, hyperextension, and occlusion gaps.
- Retarget: map the cleaned skeleton onto the target rig — MetaHuman, Unreal, Unity, Blender, Maya, or a custom skeleton — and re-check that nothing broke in the transfer.
- Hand-key the rest: markerless capture rarely nails fine fingers or subtle facial nuance, so those are keyed by hand.
- QC and deliver: review at speed and in-engine, then export FBX, BVH, USD, or engine-ready.
For more on running this remotely from your own footage, see our remote AI mocap cleanup guide.
DIY vs outsource
If you have one or two simple clips and time to learn, DIY cleanup is reasonable — the foot-lock and filter steps above get you most of the way. Outsourcing pays off when volume, deadlines, or quality climb: cleanup typically takes 2x to 8x the length of the capture, and the judgment calls (how much to filter, when to hand-key, how to fix occlusion) are where experience saves hours per clip. Many studios capture themselves and hand us only the cleanup and retarget. Either way, the work in this guide is the core of our markerless motion capture service.
Frequently asked questions
Why does my AI mocap have foot slide?
Foot-slide happens because a markerless solver guesses depth from a flat image and cannot perfectly tell when and where a foot is planted. The fix is to detect the contact frames, lock the foot in world space during those frames, and use IK to keep the leg attached while the hips move over it. Done correctly the character pushes off the ground instead of gliding.
Can I fix mocap jitter by just smoothing it?
Light, targeted smoothing helps, but heavy smoothing is a mistake — it flattens the snap, impacts, and weight that make the motion read as real. The better approach is a gentle per-joint filter aimed at the noisy frequency band while protecting the key extreme poses, so you remove the shimmer without removing the performance.
What causes knee and elbow pops in AI capture?
Pops come from the solver flipping a joint past its natural range or guessing a position through occlusion, such as when one limb hides another. They are fixed by finding the snapping frames, constraining the joint to a sane range, correcting the rotation axis so the bend points the right way, and blending the correction into the surrounding motion.
Should I clean up mocap myself or outsource it?
For one or two simple clips with time to learn, DIY is reasonable. Outsource when volume, deadlines, or quality rise, since cleanup typically takes 2 to 8 times the capture length and the judgment calls around filtering, hand-keying, and occlusion are where experience saves real time. Many studios capture themselves and outsource only the cleanup and retarget.
Bad-looking AI mocap is almost always fixable — it just needs the right cleanup passes in the right order. Send us your raw solve and your target rig, and we will return clean, engine-ready animation with the slide, jitter, and pops gone. Get a free quote.
