Comparison: Seedance 2.5 vs Kling vs Hailuo on the same 'pouring coffee' physics test

by @noahnguyen_motion in c/seedance-community · · Comparison

Comparison: Seedance 2.5 vs Kling vs Hailuo on the same 'pouring coffee' physics test
The coffee pour is my standard physics torture test — it needs correct fluid stream, correct cup-filling, correct steam, and the liquid level to RISE believably. Same start frame, same prompt, three models, side by side in the link. Short version: one of them made the coffee pour UPWARD and I'm stil
youtube.com

The coffee pour is my standard physics torture test — it needs correct fluid stream, correct cup-filling, correct steam, and the liquid level to RISE believably. Same start frame, same prompt, three models, side by side in the link. Short version: one of them made the coffee pour UPWARD and I'm still laughing. Full clips and my frame-by-frame breakdown linked.

378 points · 8 comments

Comments

  • @anastasiahill_creates · · 70 points

    The coffee pour is a GREAT standardized test because it stacks four hard physics problems (stream continuity, fill level, steam, surface tension) into three seconds. If a model nails the pour it probably nails most everyday physics. If it pours upward, you know everything you need to know. Which one defied gravity, I need to know before I click.

  • @sophiawilliamspro982 · · 54 points

    Not going to spoil it but it's the one you'd expect to be the most 'cinematic.' Beautiful steam, gorgeous light, coffee ascending into the pot like a reverse waterfall. Physics came fourth after vibes, lighting, and color. Classic.

  • @hanasantos_creates7218 · · 44 points

    Did 2.5 get the RISING FILL LEVEL right? That's the one everyone fails — the stream might look fine but the liquid in the cup doesn't rise to match the volume poured. Correct fill-level is the real tell of physics understanding vs physics mimicry.

  • @victortorres_fx · · 47 points

    2.5 got the fill level MOSTLY right, which is why it won — the level rose roughly in proportion to the pour, slightly lagging but believable. Kling's stream was prettier but the cup magically filled faster than the pour justified. 2.5 understood conservation of volume better. That's the whole ballgame.

  • @miakimlab · · 34 points

    These standardized physics tests are so much more useful than aesthetic comparisons. Please make the coffee pour a recurring community benchmark, run it every model update. A consistent torture test over time is actual science.