Tried Move AI markerless mocap from a single phone cam. First impressions after a weekend of flailing in my living room.

by @sakurahernandezcreative678 in c/dance-videos · · Mocap

Been curious whether single-camera markerless is good enough to drive AI dance yet, vs the Viggle/ref-to-video route. Spent Saturday dancing badly in front of my phone. Report:

The good:
Setup is genuinely 'prop phone up, hit record'. The onboarding is smoother than anything in ComfyUI will ever be.
Clean, well-lit, front-facing movement tracks shockingly well. Basic grooves, arm waves, steps — all solid.
Exports a usable skeleton I can retarget onto a character and then feed as controlnet. That part rules.

The bad:
The SECOND you turn more than ~45 degrees it guesses your back and it guesses WRONG. Depth ambiguity from one camera is unsolved, no shock.
Fast footwork gets smoothed into mush.
It invented a third knee at one point. I have the clip. It's going in the bloopers sub.

Verdict: for frontal, mid-tempo choreography it's a real shortcut. For spins and floorwork, stick to multi-angle reference or hand-pose it. Happy to answer setup Qs.

336 points · 7 comments

Comments

  • @henryyoung_fx581 · · 67 points

    need the third knee clip. for science.

  • @oscarcosta_studio · · 55 points

    it's posted in bloopers now, titled 'patella abundance'. you're welcome

  • @kaichenlab6445 · · 62 points

    the foot contact is cleaner than most of what gets posted here, and that's genuinely the hardest part of the whole medium. good planting on the floor.

  • @elenasingh_art · · 91 points

    single-cam markerless hitting the depth-ambiguity wall is as old as time. the human brain has two eyes and a lifetime of priors; the phone has neither. frontal-only is the honest use case and you nailed describing it.

  • @oliviajohnson_studio · · 44 points

    two cameras at 90 degrees fixes like 80% of the back-guessing if you're willing to sync them. the free apps won't do it but the paid tier does.