I compiled 300 camera-movement phrases and tagged which actually work per model

by @chenpark in c/tutorials · · Resource

I compiled 300 camera-movement phrases and tagged which actually work per model
Spent a month testing camera prompts across Kling, Hailuo, Runway and Seedance. Each phrase tagged works / ignored / unreliable per model. It is a free spreadsheet on GitHub.
github.com

Spent a month testing camera prompts across Kling, Hailuo, Runway and Seedance. Each phrase tagged works / ignored / unreliable per model. It is a free spreadsheet on GitHub.

497 points · 7 comments

Comments

  • @isabellahernandez_films376 · · 97 points

    this is the resource this whole hobby needed, 'dolly in' works everywhere but half the fancy phrases do literally nothing and we have all been cargo-culting them

  • @leomartinez_films934 · · 56 points

    the number of people who write 'dutch angle crash zoom whip pan' and get a static shot because the model ignores all three, this spreadsheet is a public service

  • @avacosta · · 47 points

    the per-model breakdown is the key part, 'orbit' being reliable in Hailuo but unreliable in Seedance is exactly the kind of thing you cannot know without testing it 50 times

  • @noahtanakaworks2048 · · 21 points

    minor methodology question: how many gens per phrase before you tagged it? small sample on a stochastic model could mislabel an 'unreliable' as 'ignored'

  • @lunatanaka_ai2234 · · 25 points

    fair, 10 gens per phrase per model, I tagged 'unreliable' when it worked 3-7/10 and 'ignored' when 0-2/10, it is in the readme, not perfect but it is 12,000 generations of data