Questo articolo non è ancora disponibile nella tua lingua; mostriamo l'originale.
Blog

Tutorial

Muse Video 1.0 Practical Guide: 5 High-Value Video Generation Use Cases

Text-to-video and image-to-video patterns, prompting notes, and cost planning for developers

Nov 16, 2026 12 min read
Abstract mint-green halftone dot field on near-black, density graduating across the frame; cover art for a tutorial.
In questa pagina
A cinematic 16:9 hero shot generated with Muse Video 1.0 — a slow dolly-in that demonstrates single-take camera control.

What Muse Video 1.0 Actually Gives You

Muse Video 1.0 is a video generation family with two routes, and understanding what is fixed versus what you control is the fastest way to scope a build. The fixed part is the output: every video is 10 seconds at 720p with built-in sound. You do not configure duration, resolution, or an audio toggle. If your product needs clips of varying length, plan to compose or trim them downstream.

The variable part is the input. The text-to-video route takes a single required prompt and an optional aspect_ratio. The image-to-video route takes a required prompt plus a required image_urls array of 1–9 reference images (http or https URLs), and an optional aspect_ratio that additionally supports auto.

Billing is per video, not per second or per token. That single unit matters more than it looks: it means your cost model is a clip counter. Ten variants of the same idea cost ten units, regardless of how long the prompt is (up to 5,000 characters) or how many reference images you attach on the image route.

Both routes sit behind one interface on HiAPI.ai, so switching routes is a parameter change rather than a new integration:

Before picking a use case, internalize the two constants — 10 seconds, 720p, with sound — and the two levers — prompt and aspect ratio. Everything that follows in this guide is a variation on those.

Text-to-Video or Image-to-Video: Choosing the Right Route

The decision rule is short: if you have no visual assets and need speed, use text-to-video; if you have brand or product assets and need fidelity, use image-to-video.

Text-to-video is the lower-friction route. You send a prompt describing subject, action, camera movement, and sound, and you get a clip back. There is nothing to prepare, which makes it the right choice for concept exploration, internal drafts, and any pipeline where the input is a sentence rather than a file.

Image-to-video is the fidelity route. You supply 1–9 reference images alongside a prompt that describes how the subject moves, how the camera moves, and the sound you want. Two practical points about image_urls deserve emphasis:

  • They guide, they do not clone. The reference images are not guaranteed to become a pixel-exact first frame. Treat them as conditioning for look, subject, and identity, not as frame zero.
  • aspect_ratio defaults to auto on this route, following the reference images. If your images are mixed shapes, set a fixed ratio explicitly.

A useful tiebreaker: ask whether a reviewer would reject the output for looking like the wrong thing. If yes, bring reference images. If the clip only needs to communicate a direction, text-to-video is enough and faster to iterate.

There is also a hybrid workflow worth knowing. Run text-to-video until a concept survives review, then, if you need brand accuracy, capture a frame of the approved concept as a reference image and re-run on the image route. You get exploration speed early and fidelity late, without changing your integration.

Use Case 1: Rapid Concept and Storyboard Previews

The highest-leverage use of text-to-video is turning written direction into motion before anyone commits to production. A storyboard describes a shot; a 10-second clip shows whether the shot reads.

The workflow

Write the shot as you would brief a director — subject, action, camera, sound — and generate one clip per beat. Because there is no asset preparation, a creative team can turn a script's key moments into moving previews in a single sitting, then review them together.

Prompt structure that works

The prompt field accepts up to 5,000 characters, but short and structural beats long and literary. A reliable pattern:

[Subject + appearance] + [action over the 10 seconds] + [camera behavior] + [sound]

For example: "A courier in a rain-soaked jacket steps off a tram and scans the street; slow push-in from behind, handheld; low city hum, distant tram bell."

Notes for this use case

  • Write motion, not adjectives. "Slow dolly forward" gives the model something to do; "cinematic" does not.
  • Respect the 10-second envelope. One action arc per clip. If a scene needs two beats, generate two clips.
  • Pick the canvas per review format. Storyboard reviews often happen on phones — set aspect_ratio to 9:16 rather than leaving the 16:9 default.
  • Budget is a clip count. A 12-beat storyboard is 12 videos. That predictability makes it easy to ask for approval on scope before generating.

Reviewers respond to motion far more concretely than to prose, and a rejected clip costs the same as an approved one — so generate the ambiguous beats first and the obvious ones last.

Use Case 2: Product and Brand-Accurate Clips from Reference Images

When the clip must show your product or your spokesperson, image-to-video is the route. You provide 1–9 reference images of the subject and a prompt describing motion, camera, and sound; the model uses the images to steer appearance.

Why reference images help

Text-to-video has to invent the look of your product. Image-to-video does not: the image_urls array carries that information in. This is the difference between a generic skincare bottle and something a brand manager will sign off on.

The first-frame caveat

The official parameter description is explicit: reference images guide the video and are not guaranteed to become a pixel-exact first frame. Plan for this in two ways:

  • Do not build a pipeline that assumes frame one equals your reference image. If you need an exact opening frame, treat the generated clip as source material to be trimmed or composited.
  • Audit a few outputs per subject before scaling. Some subjects lock in more tightly than others.

Image preparation

  • Use http or https URLs reachable by the service; host them somewhere stable rather than generating signed URLs that expire mid-batch.
  • Prefer a small, consistent set — a few clear angles of the same subject — over nine loosely related images, which can pull the output in different directions.
  • Set aspect_ratio explicitly. On this route the default is auto, which follows the reference images; if your references are mixed, auto may not match your target channel.

Prompt structure

The prompt should describe transformation, not appearance — the images already carry appearance. "The bottle rotates slowly on a marble surface as light sweeps across the label; static camera, shallow depth of field; soft ambient hum." Keep it under the 5,000-character limit and keep it about movement.

Use Case 3: Multi-Format Social Cutdowns by Aspect Ratio

One creative idea rarely ships in one shape. Both routes accept an aspect_ratio parameter with a shared enum — 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3 — and the image route adds auto.

The rule: never rely on defaults in production

Text-to-video defaults to 16:9; image-to-video defaults to auto. Both are reasonable for a first experiment and both are dangerous in a pipeline, because a silent default produces a landscape clip for a vertical placement. Always pass aspect_ratio explicitly, derived from the destination, not from the model.

A practical mapping:

  • 9:16 — short-form vertical feeds and full-screen mobile.
  • 1:1 — square placements and grid-based feeds.
  • 16:9 — landscape embeds, site heroes, presentation decks.
  • 4:3 / 3:4 — mid-ratio formats where the subject is taller or wider than square but not extreme.
  • 3:2 / 2:3 — photographic ratios for editorial layouts.

Generating a format set

Because duration is fixed at 10 seconds, aspect ratio is the only output dimension you vary per channel. Generate the same prompt family once per target ratio and you have a coherent set. Two cautions:

  • Composition changes with the frame. A prompt that fills a 16:9 frame may crop awkwardly at 9:16. Add a framing hint to the prompt — "centered subject, generous headroom" — or accept that extreme ratios need a prompt variation.
  • auto is a convenience, not a policy. On image-to-video, auto inherits the references' shape. Use it for quick drafts; pin the enum for anything scheduled.

When the same creative goes to three placements, that is three videos — a fact worth writing into your planning sheet rather than discovering in the invoice.

Use Case 4: Short Loops and Ambient Clips with Sound

Every Muse Video 1.0 clip is 10 seconds at 720p with built-in sound. That combination is unusually well suited to ambient and loop-oriented content, where the bar is atmosphere rather than narrative.

Where this fits

  • Background visuals for landing pages, menus, kiosks, or event screens.
  • Ambient filler between longer pieces in a video edit.
  • Short social posts that need presence without a script.

Prompting for ambience

Ambient clips fail when the prompt asks for a story. Ask for a state instead:

  • Describe sustained motion: drifting fog, rippling water, a slow light sweep, traffic blur.
  • Specify the camera as steady — "locked-off tripod shot" — so the frame can be looped, or describe a single continuous move if a loop is not required.
  • Name the sound explicitly. The prompt field is where you describe the sound you want; built-in audio does not mean audio you can leave unspecified. "Distant rain on glass, faint low drone" produces something far more usable than a generic request for "ambient audio."

Handling the loop seam

A 10-second clip is not automatically seamless. Two practical approaches:

  1. Request motion that reads as continuous — drifting, pulsing, flowing — so a hard cut at the seam is less noticeable.
  2. Trim and cross-fade in your editor rather than relying on the model to close the loop.

Aspect ratio, again

Ambient content is often displayed vertically (kiosks, phone backgrounds) or as a wide banner. Pass aspect_ratio to match the surface; a 16:9 default in a 9:16 slot will need cropping and may lose the motion you asked for at the frame edges.

Because each clip is a single billable unit, an ambient library is a matter of count: decide how many distinct atmospheres you need, generate them, and keep the winners.

Use Case 5: Batch-Generated Variants for Testing

Per-video billing is what makes creative testing tractable. Your spend scales linearly with the number of clips, so the question "how many variants can we afford to test?" has a direct answer: as many as you decide to generate.

Why variant testing suits this model

Creative performance is empirical. Instead of debating which prompt framing or reference set will land, generate a batch and measure. Because duration is fixed at 10 seconds and resolution at 720p, variants are directly comparable — the only differences are the ones you introduced.

Designing a variant batch

Change one dimension at a time so results are attributable:

  • Prompt angle — same subject and camera, different opening beat.
  • Aspect ratio — same creative across 9:16, 1:1, and 16:9 for different placements.
  • Reference set — on the image route, vary which 1–9 images you supply, or their order and mix.
  • Route — text-only versus image-guided for the same concept, to see whether references change the read.

Operational notes

  • Hash your inputs. Store prompt, route, aspect ratio, and reference URLs with each output so a winning variant can be reproduced exactly.
  • Keep prompts under 5,000 characters and keep them clean; batch runs are a poor place for untested prose.
  • Watch the seam between batches. If you add images later, remember auto aspect ratio will follow them.
  • Batch size is your budget lever. Ten variants is ten videos; one hundred is one hundred. There is no per-second drift to model.

This is also the cheapest way to learn a new model's behavior: a small structured batch tells you more about how Muse Video 1.0 interprets motion and sound language than any amount of prompt theorizing.

Prompt, Parameter, and Cost Notes for Production

A consolidated checklist before you wire Muse Video 1.0 into anything that runs unattended.

Parameters

  • prompt — required on both routes, min length 1, max length 5,000. Describe subject, action, camera movement, and sound. On the image route, describe how the referenced subject moves rather than what it looks like.
  • image_urls — required on image-to-video, 1–9 http/https URLs. Fewer, clearer references usually beat a loose set of nine.
  • aspect_ratio — text route: 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3, default 16:9. Image route: the same enum plus auto, default auto. Set it explicitly in production.

Constants to design around

Ten seconds. 720p. Built-in sound. None of these are parameters. If a brief needs 30 seconds, that is three clips and an edit; if it needs a resolution not offered by the model, that is a downstream step.

Cost planning

The billing unit is per video, at a single tier. That gives you a clean formula: videos generated × the per-video price = spend. Budgeting by clip count beats budgeting by seconds, because clip count is what your pipeline actually requests.

Two habits that keep that number honest:

  • Count retries. A rejected clip costs the same as an approved one. A storyboard of 12 beats that takes two attempts each is 24 videos, not 12. Plan for a retry factor and track it.
  • Count ratios separately. Three aspect ratios for one creative is three videos, not one.

The official rate is $0.0336 per clip on both routes, for a fixed 10-second 720p video with built-in sound. Confirm the live price panel if you need to re-check, and budget against this official figure.

The two route pages

Start with the route that matches your inputs, pin your aspect ratios, and budget in clip counts — that is enough to ship a first version of any of the five use cases above.

Inizia a creare con artefatti IA persistenti

Ottieni la tua chiave API e chiama modelli di immagini, video e audio tramite un'unica API basata sui task.

Leggi la documentazione API