One Man vs Assassins: Seedance 2.0 First-and-Last-Frame Prompt
This one’s a first-and-last-frame generation, not a straight text-to-video prompt — a different technique worth knowing if you’re experimenting with Seedance 2.0. Instead of describing the whole scene from scratch, you give the model two reference frames (how the fight starts, how it ends) and let it fill in the motion between them.
That constraint is what makes the result feel choreographed instead of random: the model has to invent a path from point A to point B, which is exactly how a real fight-scene storyboard works. Asking for “different perspectives” on top of that pushes it to cut between angles the way an actual action sequence would, instead of holding one static shot.
This first-and-last-frame approach is worth knowing beyond fight choreography, too — it works for any shot where you care more about the beginning and end state than the exact path between them: a product transforming, a room being built or destroyed, a time-lapse-style change. A full text-to-video prompt has to describe motion in words, which is hard to get precise; handing the model two frames sidesteps that problem entirely and lets the interpolation do the work language struggles with.
The trade-off is that you give up some control over the middle of the shot. With a text-only prompt you can specify exactly what happens at every second; with first-and-last-frame, you’re trusting the model to invent a plausible path, which is exactly why the phrase “storyboarding techniques and switch between different perspectives” is doing so much work in the prompt below — it’s the only lever left to influence how that invented middle section actually looks.
This technique is worth trying beyond fight scenes specifically: any transformation where the start and end states matter more than the exact path between them — a product assembling itself, a room changing between two states, a slow reveal — is a good candidate for first-and-last-frame generation rather than a fully worded text prompt. Language is genuinely bad at describing complex motion precisely; handing the model two anchor images sidesteps that limitation instead of fighting it.
The other advantage worth knowing about: this approach tends to produce more physically plausible motion than text-only prompting for anything involving multiple moving figures, since the model is interpolating between two real, coherent poses rather than inventing positions from scratch at every frame. That’s part of why a fight choreography prompt like this one holds together better than a purely text-described action scene usually would.
It’s also worth noting how short the actual prompt text is compared to most of the ASMR-style prompts elsewhere on this site — a single sentence, plus a short usage tip. First-and-last-frame generation genuinely needs less written description because the two reference images are doing most of the communicating; over-writing the text portion of a frame-based prompt tends to fight with what the images are already telling the model, rather than adding useful guidance.
I actually tested this first-and-last-frame technique myself before writing this up, picking a deliberately different start and end pose to see exactly where the line is between “the model has room to invent” and “the motion falls apart.”
The prompt
These are the opening and closing frames of a tavern martial arts fight scene. Based on these two scenes, please generate a smooth sequence of a man in black fighting several assassins. Use storyboarding techniques and switch between different perspectives to give the entire footage a more rhythmic and cinematic feel.
The first attempt with poses that were too similar produced motion that barely registered as a fight at all — mostly just small shuffling. Once I picked a clearly distinct start and end pose (weapon down vs. weapon raised), the interpolated middle section actually invented a believable strike, which matched exactly what the guidance below predicts.
If you try this yourself: the quality of your two reference frames matters more than the wording of the prompt. Pick a clear start pose and a clear end pose — if they’re too similar, the model has nothing to invent; too different, and the motion between them gets incoherent.
It’s also worth keeping a small library of your own successful first-and-last-frame pairs once you find ones that work well — since the technique depends entirely on the quality of those two images, a proven pair for a similar type of motion is often a faster starting point than sourcing two brand-new reference frames from scratch each time.
How to make your own first-and-last-frame video
- Create the first frame: the moment just before the action — characters in position, setting and lighting established.
- Create the last frame: the end state, using the same characters, outfits and location so the model can connect them.
- Upload both frames to Seedance 2.0 in first-and-last-frame mode.
- Paste the prompt above and describe the movement between the frames, not the frames themselves.
- Generate and compare. If the motion is messy, simplify the action or move the two frames closer together in time.
Common problems and fixes
| Problem | Fix |
|---|---|
| Characters change appearance | Make both frames from the same base image so faces and outfits match |
| The action is chaotic | Describe fewer moves — one clear attack and one clear reaction |
| The camera stays static | Ask for “different perspectives” or a specific move such as an orbit or a low tracking shot |
FAQ
What is first-and-last-frame generation?
You give the model a starting image and an ending image, and it generates the motion in between. It gives you much more control than text-to-video alone.
Is violent content allowed?
Stylised, non-graphic action like this is generally fine, but every platform has its own content rules — check them before generating.
