Motion Test: cloth, hair, water, smoke — the four horsemen, ranked by how badly 2.5 handles them

by @elenatanaka_films in c/seedance-community · · Motion Test

Ran a controlled test of the four classic 'hard motion' categories on 2.5, same settings, and ranked them from most-solved to still-cursed. Your mileage will vary but the ORDER has been consistent across dozens of my gens.

SMOKE (most solved). Honestly great now. Smoke and vapor are forgiving because they're SUPPOSED to be formless and swirly, so the model's natural tendency toward soft chaos is actually CORRECT for smoke. The one place 'AI softness' is a feature. Looks genuinely convincing.

WATER (good, with caveats). Flowing and pouring water is strong, surface interaction and droplets are weaker. Big volumes = good, fine detail (splashes, individual drops) = iffy. Conservation of volume mostly holds (see the coffee test).

HAIR (inconsistent). Long flowing hair in wind can be gorgeous OR can turn into a writhing mass of tentacles with no warning. High variance. When it's good it's secondary-motion heaven, when it's bad it's body horror. Keep hair shorter/simpler to reduce the risk.

CLOTH (still cursed). Loose, flowing fabric is the final boss. Capes, long skirts, flags — the model loses track of which part of the cloth is which, and you get self-intersecting, teleporting fabric that folds through itself. Fitted clothing is fine; FLOWING cloth is where 2.5 still dies. Avoid big flowing garments on moving subjects if you want reliability.

Summary: lean into smoke, trust water for big volumes, gamble on hair, and fear the cape.

349 points · 7 comments

Comments

  • @ravimiller_ai · · 44 points

    'Fear the cape' should be the sub's motto. Flowing cloth is genuinely the hardest and your diagnosis of WHY (the model loses track of which part is which, so it self-intersects) is exactly right. Cloth needs persistent part-identity across frames and the model has none. Fitted clothes, always, if you can.

  • @rileyrobinsonlab · · 27 points

    The self-intersecting teleporting cape is a cursed sight we've all generated. 'Which fold is which' is a correspondence problem the model doesn't solve, so over time the fabric scrambles. Short, fitted, or heavily-moving-so-the-chaos-reads-as-motion are your only options.

  • @nikolaimartinez_films · · 60 points

    'The one place AI softness is a feature' for smoke is such a sharp observation. The model's weakness (formless, drifty, can't hold hard detail) is EXACTLY what smoke is supposed to do, so its failure mode and smoke's correct behavior coincide. Play to that. Prompt more smoke, atmosphere, vapor, dust — it's free realism in a category the model can't get wrong.

  • @sakurasantosworks · · 44 points

    This is why atmospheric clips (foggy, smoky, dusty, steamy) so reliably look GOOD from these models — the whole scene is made of the stuff the model can't mess up because it's already supposed to be formless. 'Set it in fog' is a cheat code for believability.

  • @oscarwilliamscreative7485 · · 29 points

    This ranking is so useful as a SHOT-DESIGN guide. Now I know to build atmospheric smoky scenes (easy win), keep my characters in fitted clothes (avoid the cape), and either commit to great hair or keep it tied back. Designing TO the model's strengths. Great post.

  • @felixmartinez_vfx · · 35 points

    Hair being 'secondary-motion heaven or body horror with no warning' is painfully accurate. The variance is the worst part — you can't PLAN around it, you just roll the dice and sometimes get Medusa. Shorter hair genuinely reduces the tentacle risk, can confirm.