GPT Image 2From $0.007/imageFive field-tested text-to-video prompts — camera moves, action tracking, ink-wash style, vertical 9:16, and a 12-second one-shot — each shown with the actual clip it generated on the live API.
Choose a model, enter your prompt, and see the result.
HiAPI Blog
HiAPI
Generate it with HiAPI
duration is an integer from 3 to 15 seconds, resolution is 720p or 1080p, aspect_ratio accepts 16:9, 9:16, or 1:1. The schema is strict — unknown fields like seed or negative_prompt are rejected with a 400, so typos fail loudly.Each recipe has three parts: a one-line goal, the exact prompt (copy it as-is), and the clip it generated. I wrote every prompt myself and ran it through kling-3.0-turbo on the live hiapi task API; the videos are hosted here permanently, so nothing expires or gets swapped for a cherry-picked demo. Settings are in each caption so you can reproduce the exact call.
Working habits that paid off with this model:
Goal: one continuous rising move that ends on a wide reveal, no cuts.
Low angle on rain-slicked cobblestones in an old harbor town at dusk, then one slow continuous crane up past mooring ropes and lantern-lit fishing boats, revealing the full marina glowing under a purple and orange sky. Single unbroken rising move, no cuts, stable cinematic camera, photoreal, reflections shimmering on the wet stone.
Crane-up from wet cobblestones to a lantern-lit marina — kling-3.0-turbo/text-to-video, 5s @ 720p, 16:9. Real output, raw first take, generated for this guide.
Why it works: the shot is written as a path — start low on the cobblestones, rise past the ropes and boats, end on the marina — so the model has a beginning, middle and destination instead of a static scene description. "Single unbroken rising move, no cuts" is the insurance clause; drop it and reveals like this tend to get chopped into two shots.
Goal: fast subject motion with the camera matched to it, kept coherent.
Low tracking camera racing alongside a mountain biker tearing down a narrow forest singletrack, roots and rocks blurring past the lens, dry leaves kicked up behind the rear wheel, sunlight strobing through the canopy. Camera stays locked on the rider the whole way down, fast and kinetic, photoreal.
Tracking a mountain biker down a forest singletrack — kling-3.0-turbo/text-to-video, 5s @ 720p, 16:9. Real output, raw first take, generated for this guide.
Why it works: the rider's motion and the camera's motion are described as one system ("low tracking camera racing alongside ... locked on the rider"), so the subject stays framed instead of drifting out of shot. The secondary detail — leaves kicking up, sunlight strobing through the canopy — gives the speed visible evidence without adding another subject to manage.
Goal: lock a non-photoreal style and keep it stable for the whole clip.
Traditional Chinese ink-wash animation style. A lone red-crowned crane lifts off from a misty mountain lake at dawn, slow wingbeats rippling the water, pine silhouettes dissolving into layered mist behind it. Monochrome ink gradients on visible rice-paper texture, a single red seal-stamp accent, serene unhurried motion.
Ink-wash crane lifting off a misty mountain lake — kling-3.0-turbo/text-to-video, 5s @ 720p, 16:9. Real output, raw first take, generated for this guide.
Why it works: "Traditional Chinese ink-wash animation style" is declared before any scene content, and the prompt gives the style a physical medium — ink gradients, visible rice-paper texture, a red seal-stamp accent — so the model renders a material, not a vague aesthetic. Kling handles this genre remarkably well; the output even placed the seal stamp like a real scroll painting.
Goal: a phone-native portrait clip where the performance — not the action — carries the shot.
Vertical close-up portrait of an elderly watchmaker examining a tiny brass gear through a jeweler's loupe, lit by one warm workbench lamp in a dark room. His focused frown slowly relaxes into a quiet satisfied smile as the mechanism starts ticking. Shallow depth of field, photoreal skin texture, subtle handheld feel.
Watchmaker's frown easing into a smile — vertical — kling-3.0-turbo/text-to-video, 5s @ 720p, 9:16. Real output, raw first take, generated for this guide.
Why it works: the emotional beat is written as a change over time ("focused frown slowly relaxes into a quiet satisfied smile as the mechanism starts ticking"), which gives the clip an arc to perform instead of a single expression to hold. Set aspect_ratio to 9:16 and the framing composes for the vertical canvas natively — no cropping a widescreen shot after the fact.
Goal: use the flexible 3–15s duration for a mini sequence — three story beats in one unbroken shot.
One continuous shot: the camera glides through the swinging door of a busy late-night diner kitchen, follows a waitress balancing three plates as she weaves between line cooks and drifting steam, then slides past her shoulder as the plates land on the neon-lit counter in front of a customer. Single unbroken steadicam take, no cuts, warm practical lighting, photoreal.
One-shot steadicam ride through a diner kitchen to the counter — kling-3.0-turbo/text-to-video, 12s @ 720p, 16:9. Real output, raw first take, generated for this guide.
Why it works: most text-to-video models box you into fixed 5s or 10s clips; kling-3.0-turbo takes any integer from 3 to 15, and twelve seconds is enough room for an actual sequence — through the door, follow the waitress, land on the counter. Structure the prompt as ordered beats ("glides through ... follows ... then slides past") and the model paces them across the duration you paid for.
Everything below was confirmed against the live task API before generating a single clip (invalid values return a 400 that spells out the accepted ones):
| Field | Type | Notes |
|---|---|---|
prompt | string | required |
duration | integer | 3–15 seconds; a string like "5" is rejected |
resolution | string | 720p or 1080p |
aspect_ratio | string | 16:9, 9:16, 1:1 |
The schema is strict: unknown fields (seed, negative_prompt, ...) are rejected with additional properties ... not allowed, so typos fail loudly instead of being silently ignored.
All five recipes use the same async flow — create a task, poll until it's terminal, download the mp4:
import requests, time
API = "https://api.hiapi.ai/v1/tasks"
HEADERS = {"Authorization": "Bearer YOUR_HIAPI_KEY", "Content-Type": "application/json"}
def run(prompt, duration=5, resolution="720p", aspect_ratio="16:9"):
body = {"model": "kling-3.0-turbo/text-to-video",
"input": {"prompt": prompt, "duration": duration,
"resolution": resolution, "aspect_ratio": aspect_ratio}}
task_id = requests.post(API, json=body, headers=HEADERS, timeout=60).json()["data"]["taskId"]
while True:
t = requests.get(f"{API}/{task_id}", headers=HEADERS, timeout=30).json()["data"]
if t["status"] == "success":
return t["output"][0]["url"] # signed link, expires — download immediately
if t["status"] == "fail":
raise RuntimeError(t.get("error"))
time.sleep(5)
print(run("Low angle on rain-slicked cobblestones in an old harbor town at dusk ..."))
The returned output[0].url is temporary — persist the bytes right away rather than hot-linking. Full parameter reference lives in the docs, and the kling-3.0-turbo API walkthrough covers task submission, polling, and error handling step by step.
How long can clips be, and at what resolution?
Any integer duration from 3 to 15 seconds, at 720p or 1080p. The 12-second diner one-shot in Recipe 5 is a single generation, not a stitch.
Does it do vertical video?
Yes — aspect_ratio accepts 16:9, 9:16, and 1:1. Recipe 4 is a native 9:16 generation composed for the vertical frame.
Do the clips have sound?
Every clip in this guide came back with a native ambient audio track (verified on the actual files). If you need fuller generated sound design, the Kling 3.0 omni tier lists dedicated audio options.
What does a clip cost?
Billing is per second of output: $0.13/s at 720p and $0.16/s at 1080p. The four 5s clips here were $0.65 each, the 12s long take $1.56, and a 3s 720p probe is $0.39. Live numbers on the pricing page.
Turbo or Omni?
kling-3.0-turbo is the speed tier — flat $0.13/s (720p) or $0.16/s (1080p), built for volume, and every clip in this guide came back in minutes. The omni tier adds 4K output and explicit audio pricing tiers, plus image-to-video and reference-to-video variants — see the kling-3.0-omni image-to-video guide if you need to animate an existing frame. (Turbo has an image-to-video variant too: kling-3.0-turbo/image-to-video, same pricing.)
Do these prompts transfer to other models?
The structure does — continuous camera language, style-first declarations, beats over time. We keep a matching seedance-2.0-mini recipe collection if you want to compare how another model line responds to the same patterns.
All five prompts are yours to copy. Grab an API key, point the snippet above at kling-3.0-turbo, and a 3-second test clip costs $0.39 — cheaper than most image models' 4K tier.
