Sakuga punch test - 24 frames, hand-timed vs model-interpolated, side by side. Wan 2.1

by @avacostalab in c/anime-lab · · Sakuga

Sakuga punch test - 24 frames, hand-timed vs model-interpolated, side by side. Wan 2.1
A single impact punch animated two ways: 12 hand-timed keys with deliberate smears and holds vs the model interpolating smoothly between 2 poses. The hand-timed one has WEIGHT. The interpolated one is mush. Proof that timing is everything. Clip linked.
youtube.com

A single impact punch animated two ways: 12 hand-timed keys with deliberate smears and holds vs the model interpolating smoothly between 2 poses. The hand-timed one has WEIGHT. The interpolated one is mush. Proof that timing is everything. Clip linked.

707 points · 9 comments

Comments

  • @sophiamartinez · · 100 points

    THE INTERPOLATED ONE IS MUSH. This is the single most important thing about anime motion that AI people refuse to hear: smooth is the ENEMY. Sakuga is about holds, smears, and sudden snaps - the IRREGULAR timing IS the impact. A perfectly interpolated punch has no punch.

  • @oscarcosta_studio · · 84 points

    OP - smooth interpolation is literally the opposite of good 2D action. Real sakuga animators DROP frames on the fast bits and HOLD on the impact. The AI smooths exactly the frames that should be violent and holds nothing. It optimizes for the wrong thing.

  • @zaracosta_films · · 44 points

    So is the takeaway that AI video models are fundamentally bad at sakuga because they are interpolators at heart? Or is there a way to force the holds and smears?

  • @amirmehta_creates · · 52 points

    OP - you force it by generating the KEYS yourself (the extreme poses + the smear frame as its own image) and only using the model for tiny in-betweens or not at all. Treat the model as a clean-up/color tool, not a motion author. The motion has to be AUTHORED.

  • @leotanaka_art7707 · · 38 points

    hand-timed gang rise up, the model will never understand that a 2-frame smear is worth more than 20 smooth tweens