GPT Image 2From $0.030/imageThree first-take prompt recipes for wan2.7-video/text-to-video — a photoreal scene, a locked visual style, and a camera move with a payoff — each with its exact prompt, the real clip it produced, and what made it work.
Choose a model, enter your prompt, and see the result.
HiAPI Blog
HiAPI
Generate it with HiAPI
Text-to-video is the deep end of prompting. With image-to-video you hand the model a frame and only have to describe motion — we covered that playbook here. With text-to-video there is no frame: one paragraph has to carry the scene, the subject's performance, the camera, the visual style, and — because wan2.7 generates audio natively — the soundtrack too.
This guide gives you three copy-paste prompt recipes for wan2.7-video/text-to-video, Alibaba's Tongyi Wanxiang 2.7 text-to-video model, one for each job a t2v prompt has to do: a photoreal scene, a locked visual style, and a camera move with a payoff. Every clip below is the real, first-take output of the prompt printed next to it, generated through the hiapi task API — no cherry-picking across seeds, no re-rolls. All three clips together cost $1.84.
wan2.7 t2v runs on hiapi's async task endpoint, and the input is refreshingly small — three fields:
prompt — the whole movie: scene, action, camera, style, sound.resolution — an enum: 720P or 1080P, nothing else.duration — in seconds, capped at 15.No media array, no aspect-ratio knob, no negative-prompt field. Unlike the i2v sibling, there is nothing to upload — which is exactly why the prompt has to work harder.
curl -X POST https://api.hiapi.ai/v1/tasks \
-H "Authorization: Bearer $HIAPI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "wan2.7-video/text-to-video",
"input": {
"prompt": "Night street food market in a narrow alley. A cook in a white apron tosses noodles in a flaming wok...",
"resolution": "720P",
"duration": 5
}
}'
The response returns a taskId immediately; poll GET /v1/tasks/{taskId} until status is success and download output[0].url. Two things worth knowing about what comes back:
The baseline skill: make five seconds that read as footage, not as "AI video". The recipe is an ordered stack — scene, subject action, background life, camera, sound, lens style — each in one clause:
Night street food market in a narrow alley. A cook in a white apron tosses
noodles in a flaming wok; sparks and steam burst upward, glowing under a
string of red lanterns. Customers drift past as soft bokeh in the background.
The camera slowly dollies in toward the wok. Sizzling oil crackles over the
low murmur of the crowd and distant rain. Cinematic handheld feel, 35mm lens,
shallow depth of field, warm tungsten highlights against cool blue shadows.
Parameters: resolution: 720P, duration: 5 — cost $0.50.
First-take output:
Watch what the model did with each layer: the wok flame and steam animate continuously, passers-by cross the frame as out-of-focus shapes exactly where the prompt parked them ("soft bokeh in the background"), and the dolly-in is real — by second four the wok fills the frame. Turn the sound on: the sizzle sits on top of a crowd murmur, straight from the audio clause. The structural trick is that every element got exactly one clause, and the camera got exactly one instruction. Photoreal prompts fail most often from crowding — three camera ideas and a weather system in one paragraph — not from lack of detail.
The second job t2v gets hired for: stylized content where the medium is the message. The recipe is to open with the style declaration, then keep every object inside that material world — and style the audio to match:
Handcrafted paper-cutout stop-motion style. A tiny paper sailboat rides
rolling waves made of layered blue cardboard; a paper whale surfaces beside
it, spouting a burst of white paper confetti, then dives back under. The sun
is a crinkled gold-foil disc in a construction-paper sky. Slightly jittery
stop-motion cadence, soft studio light, visible paper grain and hand-cut
edges throughout. Gentle ukulele plucks over a soft ocean whoosh.
Parameters: resolution: 720P, duration: 5 — cost $0.50.
First-take output:
The style survives every frame: the waves are visibly layered cardboard sliding past each other, the whale's spout is white paper confetti rather than water, and the gold-foil sun keeps its crinkle texture while the light plays across it. The action chain executes in the order written — whale surfaces, spouts, dives back under — which is your timeline control in a five-second clip. Two rules carry this recipe: declare the style in the first sentence (style words buried mid-prompt bleed toward realism), and translate every physical effect into the medium — asking for "paper confetti" instead of "spray" is the difference between a style and a filter.
Pure camera authorship: the scene is scenery, and the shot is the content. Name the move, give it a trajectory, and — the part most prompts skip — give it an ending:
FPV drone shot racing low over a turquoise river through a narrow slot canyon
at sunrise, banking left and right with the curves of the canyon walls, then
pulling up sharply over the rim to reveal a vast desert plain glowing gold in
the morning haze. Motion blur on the near canyon walls, lens flare as the sun
crests the rim, smooth confident racing-drone motion, no cuts. Rushing wind
and a deep ambient synth swell that opens up on the reveal.
Parameters: resolution: 1080P, duration: 5 — cost $0.835.
First-take output:
The clip has an actual arc: four seconds of banking through the canyon with motion-blurred walls, then the pull-up lands the reveal — sun-blown desert to the horizon. That arc came from the prompt's shape: racing low → banking with the curves → pulling up → to reveal. A trajectory with a destination beats a list of camera adjectives every time. Note the audio direction too — "a synth swell that opens up on the reveal" ties the sound design to the camera's story beat, which is something only a native-audio model can execute from a single prompt.
Five rules, distilled from these runs and the ones that didn't make the post:
wan2.7-video/text-to-video is priced per second of output video, with the rate set by resolution:
| Resolution | Price | 5s clip | 15s clip (max) |
|---|---|---|---|
| 720P | $0.10 / second | $0.50 | $1.50 |
| 1080P | $0.167 / second | $0.835 | $2.505 |
The three clips in this post — ten seconds of 720P plus five seconds of 1080P, all with native audio — cost $1.84 total, and the whole batch was on disk within about six minutes of the first submission because tasks run in parallel server-side. Current rates are always on the pricing page.
If you already have a frame you love, wan2.7-video/image-to-video animates it at the same per-second price — the i2v recipes post is the companion to this one. For everything you can describe but haven't shot, take any prompt above, change the nouns, and post it to the task API — the wan2.7-video/text-to-video model page has the live parameter reference. A 720P first draft costs $0.50 and comes back in minutes; save 1080P for the prompt that survives it.
