原文要点

this is 9going to be a long post, but it will help you a lot if you are trying to generate ai content : Everyone's writing these essay-length prompts thinking more words = better results, i tried that as well turns out you can’t really control the output of these video models. same prompt under just a bit different scnearios generates completley differenent results. (had to learn this the hard way) After 1000+ veo3 and runway generations, here's what actually wordks as a baseline for me **The structure that works:** [SHOT TYPE] + [SUBJECT] + [ACTION] + [STYLE] + [CAMERA MOVEMENT] + [AUDIO CUES] **Real example:** Medium shot, cyberpunk hacker typing frantically, neon reflections on face, blade runner aesthetic, slow push in, Audio: mechanical keyboard clicks, distant sirens **What I learned:** 1. **Front-load the important stuff** \- Veo 3 weights early words more heavily 2. Lock down the “what” then iterate on the “How” 3. **One action per prompt** \- Multiple actions =

原文完整内容

this is 9going to be a long post, but it will help you a lot if you are trying to generate ai content : Everyone's writing these essay-length prompts thinking more words = better results, i tried that as well turns out you can’t really control the output of these video models. same prompt under just a bit different scnearios generates completley differenent results. (had to learn this the hard way)

After 1000+ veo3 and runway generations, here's what actually wordks as a baseline for me

The structure that works:

[SHOT TYPE] + [SUBJECT] + [ACTION] + [STYLE] + [CAMERA MOVEMENT] + [AUDIO CUES]

Real example:

Medium shot, cyberpunk hacker typing frantically, neon reflections on face, blade runner aesthetic, slow push in, Audio: mechanical keyboard clicks, distant sirens

What I learned:

1. Front-load the important stuff \- Veo 3 weights early words more heavily 2. Lock down the “what” then iterate on the “How” 3. One action per prompt \- Multiple actions = chaos (one action per secene) 4. Specific > Creative [\-](https://chatgpt.com/?model=auto) "Walking sadly" < "shuffling with hunched shoulders" 5. Audio cues are OP \- Most people ignore these, huge mistake (give the vide a realistic feel)

Camera movements that actually work:

  • Slow push/pull (dolly in/out)
  • Orbit around subject
  • Handheld follow
  • Static with subject movement

Avoid:

  • Complex combinations ("pan while zooming during a dolly")
  • Unmotivated movements
  • Multiple focal points

Style references that consistently deliver:

  • "Shot on \[specific camera\]"
  • "\[Director name\] style"
  • "\[Movie\] cinematography"
  • Specific color grading terms

As I said intially you can’t really control the output to a large degree you can just guide it, just have to generate bunch of variations and then choose (i found [these](https://arhaam.xyz/veo3/) guys veo3gen\[.\]app , idk how but these guys are offering veo3 70% bleow google pricing. helps me a lot with itterations )

hope this helped <3

原帖关键讨论

OwnExplanation7081 1 points

That's old news, how are you keeping any consistency in your renders through basic + + += are you using a diffusion modal

Warm-Fruit-1125 1 points

thank you

stunspot 1 points

As opposed to Sora, where the best prompting style is to choose another platform.

评论配图

ArhaamWani 1 points

glad it helped

Overall-Awareness244 1 points

Helpful tysm

来源：https://www.reddit.com/r/ChatGPT/comments/1msb3vq/the_ai_video_prompting_guide_that_actualy_worked/