Do negative prompts actually do anything in video models? I ran a blind test and I'm annoyed

by @felixtanaka_creates in c/prompt-engineering · · Negatives

Negatives flair because this needs airing out. I love negative prompts in image models. For video models I got suspicious nothing was happening, so I ran a sloppy blind test: 20 gens with my usual negative block, 20 identical prompts with the negative field EMPTY, shuffled, rated blind by a friend.

Result: my friend couldn't tell them apart above chance. For THIS model, my carefully-tuned negatives ('blurry, deformed, extra limbs, low quality') did approximately nothing measurable.

Now I'm wondering how much of my negative-prompt ritual is just a comfort blanket. Has anyone done a rigorous test where negatives clearly helped in a video model? Genuinely asking, not trying to dunk.

228 points · 9 comments

Comments

  • @ethanlee_vfx · · 94 points

    The comfort-blanket framing is painfully real. I have a 40-token negative block I paste everywhere and I've never once tested if it helps. This post is attacking me personally and I deserve it.

  • @emmajackson_films335 · · 60 points

    The 40-token inherited negative block is the universal cope. We all have one. We all never test it.

  • @zarachen_art8815 · · 99 points

    Your test matches the dirty secret: a lot of video models either don't expose a real negative channel or weight it so weakly it's noise. The 'negative field' in some UIs is barely wired to anything.

  • @ariamiller_creates292 · · 88 points

    Where negatives DO work in video, it's usually specific CONCRETE nouns ('no text, no watermark, no subtitles'), not vague quality words ('blurry, bad, deformed'). Try the concrete-noun version in your test.

  • @mayasanchez · · 67 points

    'Blurry, deformed, low quality' is 2022 image-model muscle memory. Those words were trained into negative embeddings in a specific lineage of models. A fresh video model has no such ritual baked in.

  • @emmasmithcreative · · 50 points

    Rigorous-ish: in ONE video model I found that adding a negative actually made things WORSE because it was being concatenated into the positive context. Negatives aren't always subtractive. Terrifying.