Seedance 2.0 One-Shot Video: Anatomy of a 4K POV Flythrough Prompt

The clip below is one continuous shot: a first-person dive through a drowned temple, banking past arches, statues and drifting chains, and ending on a colossal "artificial sun" waking up inside a ring of stone. It came out of Seedance 2.0 as a single text-to-video generation — no cuts, no keyframes, no editing pass — and the rumble you hear was generated with it, from the same prompt. We shipped it as a showcase on our X account; this post is the full prompt behind it, and why each piece is there.

One-shot flythroughs are where text-to-video models usually fall apart. Ask for "a camera flying through underwater ruins" and you get one of three failures: the model cuts to a new angle halfway, it shows you the thing carrying the camera, or the flight drifts because it has nowhere to go. This prompt is built to close off all three.

The full prompt

Immersive first-person gliding camera perspective, no visible drone, diver, submarine, or flying device. The camera plunges through a vast lost underwater temple buried in the abyss, racing between broken obsidian arches, bioluminescent coral, ancient sea statues, drifting chains, whale-bone pillars, glowing blue runes carved into flooded stone walls, silver fish splitting apart, bubbles, sand particles, and huge shadows moving behind the ruins. At the end, the temple corridor opens into a colossal submerged celestial core, like an artificial underwater sun trapped inside a circular stone machine, releasing golden light waves, rising bubbles, floating stone rings, and rippling shockwaves through the water. Continuous POV shot, no cuts, aggressive banking, near-miss foreground objects, fast forward momentum, cinematic motion blur, volumetric underwater mist, caustic light rays, deep blue and gold contrast, high-end fantasy VFX, epic mysterious atmosphere, ultra cinematic, premium AI video showcase quality, 4K high definition. Audio: deep underwater rumble, rushing water pressure, bubbles streaking past, distant whale-like resonance, stone machinery awakening, low cinematic impact hits, golden energy pulse at the final reveal. No readable text, no humans, no logos, no cartoon style, no shaky low-quality camera.

It reads long, but every clause is working — Seedance 2.0 accepts prompts up to 20,000 characters, so length is not the constraint. Structure is. Here is the same prompt taken apart into the six jobs it does.

1. Lock the camera, ban the carrier

Immersive first-person gliding camera perspective, no visible drone, diver, submarine, or flying device.

The first sentence is entirely about the camera, before a single scene element appears. Models trained on real footage associate a moving underwater camera with whatever carries it, so left alone they will often render the drone, a diver's hands, or a submarine viewport into the shot. Saying "first-person" is not enough — you have to ban each plausible carrier by name. The prompt's last line does the same job for camera quality ("no shaky low-quality camera").

2. Build a corridor of near-misses

...racing between broken obsidian arches, bioluminescent coral, ancient sea statues, drifting chains, whale-bone pillars, glowing blue runes carved into flooded stone walls, silver fish splitting apart, bubbles, sand particles, and huge shadows moving behind the ruins.

A one-shot needs things to fly past. Vague scenery ("beautiful ancient ruins") gives the camera nothing to move relative to, and the result reads as slow drift even at speed. This list is deliberately dense — nine kinds of objects at different scales, from sand particles to whale-bone pillars — and the phrase "near-miss foreground objects, fast forward momentum, aggressive banking" turns that density into parallax. Foreground close calls are what sell speed.

3. Script the ending

At the end, the temple corridor opens into a colossal submerged celestial core, like an artificial underwater sun trapped inside a circular stone machine, releasing golden light waves, rising bubbles, floating stone rings, and rippling shockwaves through the water.

This is the highest-leverage line in the prompt. Without a scripted destination, a 15-second one-shot keeps going until it peters out. Give the model an explicit "at the end" event with its own detail budget — light waves, floating stone rings, shockwaves — and the clip gets an arc: journey, arrival, payoff. In the final render, Seedance paces the corridor so the reveal lands in the closing seconds, which is exactly what you want a showcase clip to do.

4. Fix the grade before the render

Continuous POV shot, no cuts... cinematic motion blur, volumetric underwater mist, caustic light rays, deep blue and gold contrast, high-end fantasy VFX, epic mysterious atmosphere...

Color and finish are cheaper to specify than to regret. "Deep blue and gold contrast" does the palette work for the whole clip; motion blur and caustic light rays are what make it read as photographed water rather than a game-engine flythrough. And "continuous POV shot, no cuts" states the one-shot constraint outright — do not assume the model infers it from "first-person."

5. Score it in the same prompt

Audio: deep underwater rumble, rushing water pressure, bubbles streaking past, distant whale-like resonance, stone machinery awakening, low cinematic impact hits, golden energy pulse at the final reveal.

Seedance 2.0 generates synced audio natively (generate_audio defaults to true). Treat the audio line like a second scene description: ambience first (rumble, water pressure), then events tied to visuals — "stone machinery awakening," "golden energy pulse at the final reveal." Because the sound events are described as part of the same moments, they land where the visual events land. The clip above uses the audio exactly as generated.

6. Close with a negative list

No readable text, no humans, no logos, no cartoon style, no shaky low-quality camera.

The negative list is the quality contract. Text and logos are the most common artifacts in ruins-and-monument scenes — models love carving gibberish into stone — and a single stray human silhouette breaks the "lost world" read. Ban what would disqualify the take, nothing more.

Run it on hiapi

One async task call, then poll:

# 1. Create the task
curl -X POST https://api.hiapi.ai/v1/tasks \
  -H "Authorization: Bearer $HIAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "seedance-2-0",
    "input": {
      "prompt": "Immersive first-person gliding camera perspective, ... (the full prompt above)",
      "aspect_ratio": "16:9",
      "duration": 15,
      "resolution": "4k",
      "generate_audio": true
    }
  }'

# 2. Poll until status = success, then download output[0].url
curl https://api.hiapi.ai/v1/tasks/$TASK_ID \
  -H "Authorization: Bearer $HIAPI_TOKEN"

Two practical notes:

• Cost scales per second, by resolution tier. As of July 2026: $0.136/s at 480p, $0.293/s at 720p, $0.729/s at 1080p, $1.486/s at 4k. This 15-second 4K run lands around $22; the same prompt at 720p is about $4.40. The workflow that follows: iterate composition at 720p, then re-run the keeper once at 4k. Current tiers are listed on the Seedance 2.0 model page.
• 4K is slow. A 15-second 4K render can take tens of minutes end to end — ours took over half an hour. Poll at a relaxed interval instead of hammering the task endpoint. For the full parameter table (first/last frames, reference images, aspect ratios), see the Seedance 2.0 API guide.

FAQ

Can Seedance 2.0 really hold one shot for 15 seconds?

Yes. Fifteen seconds is the model's maximum duration, and it will hold a single move that long if the prompt demands it ("continuous POV shot, no cuts") and gives the camera a route with a destination. Long durations without those two things tend to invite cuts.

Does the audio come from the same generation?

Yes — generate_audio defaults to true, and an Audio: line in the prompt steers the mix. Nothing in the clip above was added in post.

Should I generate at 4k directly?

Only the final take. Composition, pacing, and reveal-timing problems all show up fine at 720p, at about a fifth of the 4K per-second price. Iterate there, then render once at 4k.

What if I need longer than 15 seconds?

Seedance 2.5 pushes single-shot generation to 30 seconds — see Seedance 2.5 vs Seedance 2.0 for what else changes. For most showcase work, though, 15 seconds with a scripted reveal is plenty.

---

The prompt above is free to steal. Swap the temple for a slot canyon, a shipwreck, a server hall — keep the six jobs intact: lock the camera, ban the carrier, stack the corridor, script the ending, fix the grade, score the sound. Then run it against Seedance 2.0 on hiapi and see where your corridor ends.