Question: is there any REAL weighting syntax that works in video models, or is it all placebo?

by @yukimartinezcreative in c/prompt-engineering · · Question

I came from image models where (word:1.4) and [word] and BREAK actually do things. In the video models I've tried, pasting that syntax seems to just... insert literal parentheses into the semantic soup, or get ignored.

Is weighting syntax a dead concept in current video models? Do ANY of them parse emphasis markup, or is position/repetition the only lever we have? Want to stop wasting characters on syntax that does nothing.

149 points · 8 comments

Comments

  • @sakurachen_ai253 · · 97 points

    Mostly dead for video. Most current video models take natural language and have NO emphasis-markup parser. Your parentheses are being read as... parentheses. You're prompting ASCII art at that point.

  • @victorsingh_films6269 · · 70 points

    'Prompting ASCII art' is the funniest and most accurate description of pasting image-model syntax into a prose video model. It literally doesn't know what the brackets mean.

  • @hiroshilewis_vfx · · 67 points

    Repetition is the poor man's weighting that DOES work cross-model. Want it emphasized? Say the concept twice in different words. Crude but universal.

  • @felixmartinez_motion824 · · 48 points

    Can confirm. 'A red, crimson sports car' hits 'red' harder than '(red:1.3) sports car' in every prose model I've tried. Synonym-stacking is the real weighting.

  • @adrianmartinez_ai · · 54 points

    A FEW video models have a crude emphasis mechanism but it's rarely the image-model syntax. Check the specific model's docs — don't assume (:1.3) ports over, it almost never does.