Skip to content
Lab

commit 7be0d48 ·

Garbage in, garbage out: the driving video matters more than the prompt

The output can't be better than the performance, the reference image and the precision of the prompt.


The workflow behind the Haj sounds more complicated than it is. Every shot is three inputs:

  1. A driving video. A real person performing the line: body language, timing, mouth movement.
  2. A character image. One strong reference of the Haj.
  3. A prompt. What should change, what must not.

The output can't be better than the weakest of the three.

The driving video is the performance

If the driving video is stiff, the Haj is stiff. If the lips are lazy, the sync is lazy. The model transfers what it sees; it does not direct. Most "AI problems" in the early tests were really performance problems that the model faithfully copied.

The image is the identity

A soft, badly lit reference gives the model room to improvise, and it will: a different nose in one shot, younger skin in the next. A sharp, well-lit, front-facing reference is what keeps the character the same person across an episode.

The prompt is precision, not length

A longer prompt is not a better prompt. What matters is how exactly it describes the thing you want, and the things you don't. The useful prompts for the Haj didn't come from a template. They came from repetition: run it, see the same failure twice, describe that failure away, run it again.

A prompt gets good by fixing the same mistake until it stops happening.

In practice

  • Record the performance before touching any AI tool.
  • Spend time on the reference image; it pays back in every shot.
  • Keep a prompt log. When a failure repeats, add one precise line against it, not a paragraph.
Next notec19e5a0Higgsfield → Runway → HeyGen → Kling: why each switch happened

HAJFLIX://contact

Talk to the Haj

Pick the way you like. WhatsApp is the fastest; he answers between naps.

«عندك شغل؟ الحاج فاضي… لحدي ما يجي الزبون الجاي.»