Hey, this is Scenic. Have you tried making fight scenes with Seedance? Characters move, sure, but something feels off — the camera's static, the impacts don't land. If your prompt just says "a fighting scene," the AI has no idea how the fight actually works or where the camera should be.
In this post, we break down five battle scene prompts that actually delivered results in Seedance. From retro kung fu to post-apocalyptic boss fights — here's the language that brings each genre to life.
1. 80s Garage Kung Fu — Retro Action
See the full prompt on scenic.sh →
"A fast-paced kung fu fight in a cramped garage, 80s action cinema style, grounded and physical. One protagonist faces 3 attackers, using improvised objects (chair and ladder) with fluid movement."
This prompt works because it nails the genre reference. "80s action cinema style" — that one phrase guides the output toward raw, physical, no-CG action with that unmistakable retro texture. Seedance responds strongly to style tags, so specifying which era of action you want makes a huge difference.
Despite being short, this prompt packs in "cramped garage" (spatial constraint), "improvised objects" (improvised weapons), and "3 attackers" (outnumbered). Setting the conditions and constraints of a fight produces far more specific results than simply describing "a fight scene."
2. The Waitress Strikes Back — Urban Close-Up Action
See the full prompt on scenic.sh →
"Cinematic-quality visuals, a spacious vintage grocery store (neon lights reflected on the glass display window, warm and cold lights intertwined). Subject: A beautiful waitress, very elegant and with a determined look. Action: Two armed robbers confront the waitress, one robber is nervously looking around, pointing a gun at the waitress's head..."
The strength here is translating an emotional arc into physical action. The shift from "pretending to be scared" to a sudden counterattack is explicitly scripted, so Seedance creates action with a narrative — not just random combat. The lighting description (neon lights, warm and cold intertwined) guides the scene's atmosphere through specific color temperatures, not vague "make it cool."
The camera direction is well-designed too. "Handheld camera style, rapid shaking, focus switching" builds tension into the movement itself, and writing out the camera flow shift — "tense static shots → chaotic close-up follow shots" — lets the AI create a varied editing rhythm instead of monotonous footage.
3. Bank Heist Operative — Gun-Fu
See the full prompt on scenic.sh →
"Character A highly trained female operative in a sleek black tactical suit, calm, focused, controlled breathing, sharp eyes scanning constantly. Dual pistols in hand, precise and efficient. 0-3s Camera starts inside a luxury bank. Slow push through a tense standoff—security guards aiming, civilians frozen. She stands in the center, perfectly still..."
The timecode structure is what makes this prompt special. Breaking scenes into "0-3s", "3-5s" segments lets Seedance reflect tempo changes precisely within a single sequence. The structure — static tension building into explosive movement — gets followed faithfully instead of being freely interpreted by the AI.
The character description is noteworthy too. Instead of describing appearance, phrases like "controlled breathing, no wasted motion" describe the quality of movement — that's what creates John Wick-style restrained gun-fu. When defining a character's combat style in Seedance, describe the nature of their movement before their looks.
4. Office Brawl — One-Take Close-Quarters Combat
See the full prompt on scenic.sh →
"15-second, no-dialogue, fully immersive high-speed brutal fight short film. Two characters. Live-action realistic style. Close-quarters, high-speed chaotic melee combat with no rules, fully out-of-control brawling. No flashy choreography—pure raw. Continuous exchange of punches, kicks, blocks, body checks, throws..."
"No flashy choreography—pure raw" is the core instruction. When you want a messy, realistic fight instead of polished movie choreography, the negative-instruction approach is surprisingly powerful. Telling the AI what you don't want narrows its interpretation range more effectively than describing what you do want.
The environment setup is precise too. "Modern open-plan office, desks, computers, cabinets, no glass-breaking" includes both objects and destruction constraints. This detail turns a generic brawl into a calculated fight in an actual workspace, and "Final 2-3s: deadlock" designs the climax to end on an open-ended standoff.
5. Wasteland Warrior vs. Boss — Post-Apocalyptic Action
See the full prompt on scenic.sh →
"SUBJECTS: A female post-apocalyptic future warrior, short hair pressed backward by high-speed airflow, face marked with dust and fine scars; wearing layered wasteland tactical gear: a tight heat-resistant inner layer + a damaged outer drape with hanging straps, partially embedded ceramic armor plates with scorched edges; one arm is a mechanical prosthetic..."
The density of character description is on another level. Specifying clothing layers, materials, and damage levels produces VFX-quality results. Details like "ceramic armor plates with scorched and worn edges" cause the AI to reflect the world's temporal context — the wear tells a story.
The COLOR LOGIC section is unique to this prompt. "Overall low-saturation cold gray-blue environment; the fissure-flame spear is the only high-brightness warm color" — intentionally designing color contrast controls where the viewer's eye goes across the entire frame. In complex battle scenes, using color to guide attention dramatically elevates the final result.
Common Rules for Battle Scene Prompts
Comparing all five prompts reveals shared patterns.
First, specify your genre reference. "80s cinema style", "John Wick gun-fu", "live-action realistic" — naming the action tradition you want anchors the output's direction.
Second, segment the sequence with timecodes or stages. Describing a 15-second video as one block lets the AI interpret freely. Breaking it down by seconds or action stages ensures your intended tempo changes come through.
Third, design camera movement alongside the action. "Camera orbiting as she moves", "handheld rapid shaking" — describing camera behavior within the scene prevents static, boring shots.
Try the original prompts and generate your own at scenic.sh.
