Hair-in-wind sakuga loop: Kling holds the silhouette but kills the linework, Wan holds lines but stutters

by @marcustaylor_films in c/anime-lab · · Motion Test

Motion Test, documenting a genuine model tradeoff. Trying to get a clean looping shot of a character's long hair blowing in wind (the classic beauty-shot sakuga). Ran it through both:

Kling 2.0: gorgeous flowing SILHOUETTE motion, the hair moves like real anime hair (big confident strands, good secondary motion). BUT the linework inside the hair boils and the strand-lines dissolve. Great shape, mushy detail.

Wan 2.1: holds the LINE detail noticeably better (the individual strand lines stay crisp-ish) BUT the motion stutters and the timing is uneven, like it is skipping frames unpredictably. Good detail, bad flow.

So right now it is a genuine pick-your-poison: Kling for SHAPE, Wan for LINE, neither for both. Anyone found a model or a trick that gets both the flowing silhouette AND the stable strand-lines? Clips of both in comments.

457 points · 8 comments

Comments

  • @miacostalab · · 68 points

    This is the exact shape-vs-line tradeoff that defines anime video right now and you documented it cleanly. The dream is Kling's motion with Wan's line stability. The workaround some people use: Kling for the motion, then a per-frame lineart-ControlNet stabilization pass to re-impose crisp strand lines on Kling's gorgeous shapes.

  • @amirwatanabe_creates700 · · 51 points

    OP - the 'Kling motion + lineart re-stabilize pass' is the combo I keep hearing is theoretically the answer and keep failing to execute cleanly. The re-lined frames flicker because the line pass is per-frame independent. Temporal line consistency AGAIN, our eternal boss fight.

  • @alexeiwatanabe_art · · 41 points

    Wan's stutter can sometimes be smoothed by generating at a higher frame count and then dropping to your target with GOOD interpolation (not the naive kind). But then you risk re-introducing the mush. Everything in anime video is a tradeoff against something else.

  • @yukimiller_creates · · 33 points

    OP - yeah I tried the interp-smoothing and it brought back exactly the boil I was avoiding. You rob Peter (line) to pay Paul (motion) no matter which direction you go. There is no free lunch in this specific shot yet.

  • @sakuratanaka_art · · 38 points

    documenting a clean A/B tradeoff like this is EXACTLY what Motion Test flair is for, this is more useful than a hero clip, saving as reference for which model to reach for