Hey everyone! 👋 I've been absolutely loving Dream Machine for the past few weeks—it's such a game-changer for creating short, dynamic clips. But as someone who came from using mainly static image generators, wrapping my head around prompting for *actions* has been a bit of a learning curve.
I've had my share of... let's call them "abstract interpretations" from the AI. Think "a cat jumping" resulting in a feline slowly stretching into a weird, elongated blur. Not exactly the crisp action I envisioned!
Through a lot of trial and error (and some fantastic tips scattered in other threads), I've landed on a prompt structure that's been giving me much more consistent and intentional results. I think it could really help other newcomers.
The key seems to be being **extremely specific** about the action, the subject, the camera, and the timing. A good format to build on is:
**1. Subject & Scene:** Start with a clear description of your main character and the environment.
* Example: "A gray tabby cat on a polished wooden kitchen floor."
**2. Core Action:** This is the most important part. Use strong, specific action verbs and define the *start* and *end* state.
* Instead of "jumping," try: "The cat crouches down, then leaps powerfully into the air, landing gracefully on a nearby countertop."
**3. Camera Direction:** Tell the AI *how* to show the action. This is crucial.
* Example: "The shot is a dynamic side view, following the arc of the jump, keeping the cat in frame throughout."
**4. Style & Details:** Add any aesthetic or mood filters last.
* Example: "Cinematic lighting, soft morning sun, realistic detail, slow-motion effect during the leap."
Putting it all together, a strong prompt looks like:
> "A gray tabby cat on a polished wooden kitchen floor. The cat crouches down, then leaps powerfully into the air, landing gracefully on a nearby countertop. The shot is a dynamic side view, following the arc of the jump. Cinematic lighting, soft morning sun, realistic detail, slow-motion effect during the leap."
Some extra lessons I've learned:
* **Keep it tight:** Prompts that are too long and list every single detail sometimes confuse the model. Start with this structure and then iterate.
* **Action verbs are your friend:** "Spins," "dashes," "twirls," "shatters," "unfolds," "reaches for," "turns to face."
* **Camera terms help a lot:** "Tracking shot," "close-up on," "steady cam," "low angle," "panning left to right."
I'd love to hear from more experienced users! Do you have any other formatting tricks or must-include keywords that make actions pop? What was the breakthrough moment for your action prompts?
Clean data, happy life.
I'm a junior cloud engineer at a mid-sized e-commerce company, and we run all our batch processing workloads on a mix of EC2 and Fargate.
**Primary Use Case Fit:** It's best for automating a single, linear action within a video clip, like a jump or a turn. For complex, multi-step character animation with scene changes, it gets confused and you'll need heavy editing.
**Real Output Consistency:** In my tests, about 7 out of 10 generations match the prompt's described motion if you use their frame-by-frame guidance format. The other 3 will have timing or object distortion issues.
**Input Effort & Cost:** You need a very detailed, sequenced prompt for good results, which takes time to write. At about $0.09 per generated second, experimenting with different action verbs for a 5-second clip can cost a few dollars in credits quickly.
**Honest Limitation:** It struggles with physics. Prompting for "a book falling off a shelf" often results in a weird, slow float instead of a gravitational fall. You have to add phrases like "with sudden acceleration" to get close.
For a complete newbie just wanting to make short, fun action clips, I'd stick with Dream Machine and master their specific format. If you need frame-perfect control for a commercial project, tell us your budget and if you have a video editor on your team.
Still learning