40 shots into a music video and consistency is slowly killing me — venting + asking for help

by @noahsmithlab in c/hailuo-community · · WIP

WIP, maybe 60% done. It's a performance-driven piece, one artist, lots of close-ups. The artist looks like a slightly different person in about a third of my clips and I'm losing my mind matching them.

My current process: one master keyframe of her face, I2V from variations of it. But any time she emotes hard or turns fast, the face identity slips. Re-rolling is eating my budget alive.

How are people keeping ONE human consistent across 40+ emotional close-ups? Is there a trick beyond 'grade it and pray'? Starting to think I should've cast an AI character with simpler features.

121 points · 10 comments

Comments

  • @jordanthomas_ai57 · · 84 points

    Practical: don't I2V every close-up from scratch. Generate ONE good 10s performance, then cut multiple angles/moments out of it. One source = one consistent face.

  • @miamillerlab8209 · · 52 points

    Wait, mine the coverage out of fewer longer gens instead of many short ones. That's actually genius for a music video.

  • @liamgarciacreative9224 · · 96 points

    Simpler features is unironically the move. Strong distinctive bone structure + a signature element (hair, glasses, a scar) survives the drift. Forgettable faces melt.

  • @sashagarciacreative6587 · · 61 points

    Yeah I gave my lead a very specific undercut and a nose ring. The model clings to those landmarks even when the face wobbles.

  • @henrypatelpro230 · · 58 points

    Honestly 40 shots of a consistent human is near the edge of what's reasonable right now. You're not doing it wrong, the tech is just hard here. Hang in there.

  • @ethankimlab5320 · · 44 points

    Grade-and-pray is real but a subtle face-consistency pass in post (there are tools for this now) will unify the drift better than chasing perfect gens.