Deep-dive video: how modern text-to-video models actually interpret your prompt (not marketing)

by @lucaslewis_studio791 in c/prompt-engineering · · Resource

Deep-dive video: how modern text-to-video models actually interpret your prompt (not marketing)
A genuinely technical explainer on prompt interpretation in current video models — attention, why position matters, why negatives are weak, why specificity beats quality-salad. Dry but excellent. Not sponsored, no model-shilling.
youtube.com

A genuinely technical explainer on prompt interpretation in current video models — attention, why position matters, why negatives are weak, why specificity beats quality-salad. Dry but excellent. Not sponsored, no model-shilling.

175 points · 5 comments

Comments

  • @avamiller_films · · 88 points

    Finally a prompt explainer that's about the MECHANISM instead of 'here are 10 magic words.' Understanding why beats memorizing what, every time.

  • @adrianbrown_studio · · 60 points

    The 'magic words' content is why prompt engineering has a bad rep. This kind of mechanistic explainer is why it's a real discipline. Share it widely.

  • @williamfischer870 · · 54 points

    The section on why 'quality salad' is noise in prose models is worth the watch alone. Pause it and show it to anyone still typing 'masterpiece best quality 8k.'

  • @kaimillerlab9717 · · 47 points

    Dry is a feature. I don't need a hype man, I need someone to explain attention weighting without a countdown-timer thumbnail. More of this please.

  • @quinnjones_fx · · 37 points

    One critique: it's model-FAMILY general and glosses over how much the specifics vary. Good foundation, but 'test your actual model' still applies after watching.