Guide
Muse Video 1.0 Developer Guide: Capabilities, API Parameters, Pricing & Integration
Text-to-video vs image-to-video: route parameters, per-video pricing, and practical integration advice
На этой странице
What Muse Video 1.0 Actually Produces
Before you write a single line of integration code, fix the output contract in your head: Muse Video 1.0 produces a video. Every request that succeeds returns a clip that is 10 seconds long, 720p, and includes built-in sound. Those three properties are not parameters — there is no duration knob, no resolution enum to choose from, and no separate audio toggle. They are constants that hold across both routes in this family.
That has an immediate architectural consequence: your pipeline can assume a known, fixed artifact shape. Downstream steps — thumbnail extraction, caption generation, concatenation, CDN naming conventions, player configuration — can be written once against a 10-second 720p asset with an audio track. You are not branching on output metadata.
It also means the model is not the place to solve timing problems. If a scene needs to run 30 seconds, you are generating three clips and editing them together, not asking for a longer render. If a scene needs a different frame shape, you are choosing an aspect_ratio at request time, not post-cropping a mismatched render.
Keep two inputs in mind, because everything else in this guide follows from them: text (a prompt) and, on one of the routes, a set of reference images. Both feed the same fixed output box. The only real design decisions you have are which route you call, what you put in the prompt, and what aspect ratio you ask for.
The Two Visible Routes and When to Use Each
Muse Video 1.0 exposes two routes under the same family, and they are close to a pure split on input type.
muse-video-1.0/text-to-video
Capability: text-to-video. Required parameter: prompt only. You describe a scene in language and get a 10-second 720p clip with sound back. This is the route for generative B-roll, concept exploration, storyboard visualization, and any workflow where nothing exists yet — you are inventing the frame rather than extending one.
muse-video-1.0/image-to-video
Capability: image-to-video. Required parameters: prompt and image_urls (1–9 reference images, http or https URLs). Here the video is guided by visual references. The prompt's job shifts: instead of describing a world from scratch, you describe how the subject in the reference images moves, how the camera moves, and the sound you want.
This route fits product shots that need motion, character or mascot animation, style-consistent series where each clip must sit next to the last, and any case where you already have approved imagery and want video that stays in that visual neighborhood.
Choosing between them
Ask one question: do I already have the look? If the answer is no, start from text. If the answer is yes and the look is non-negotiable, use image-to-video and supply references. A practical hybrid: prototype on the text route to discover what motion and camera language works, then move to the image route once you have final art, reusing the prompt structure that proved itself.
Both routes are visible as separate model entities — text-to-video and image-to-video — so you can wire them as distinct model references in your code rather than branching on a flag.
Key Parameters: prompt, aspect_ratio, and image_urls
The parameter surface is intentionally small. Here is the full set, per route, exactly as documented.
prompt (both routes, required)
Type: string. Minimum length 1, maximum length 5000 characters. Required on both routes. The documented intent is to describe the subject, action, camera movement and sound. On the image-to-video route the wording is narrower: describe how the subject in the reference images moves, how the camera moves, and the sound you want.
Treat 5000 characters as generous headroom, not a target. Validate it client-side and reject empty strings before you spend a request.
aspect_ratio (both routes, optional with different defaults)
This is the parameter most likely to trip you up, because the enum differs between routes.
- text-to-video:
16:9,9:16,1:1,4:3,3:4,3:2,2:3. Default16:9. Note thatautois not in this list. - image-to-video: the same seven fixed values plus
auto. Defaultauto, which follows the reference images. You can also force a fixed ratio if you need one.
image_urls (image-to-video only, required)
Type: array. Minimum 1 item, maximum 9 items. Each entry is an http or https URL pointing at a reference image that guides the video. The documentation is explicit that these references are not guaranteed to become a pixel-exact first frame — they steer content and style, they do not lock frame one.
Nothing else
There is no duration parameter, no resolution parameter, no seed, no negative prompt, and no audio toggle in the documented schema. Do not send fields you invented; build your request object from the sets above and keep duration and resolution as assumptions in your downstream code.
Pricing Model: Flat Per-Video Cost
Billing on both routes is per video, at a single tier: $0.0336 per video. There is no size-based or duration-based variation, because there is no size or duration to vary — every call returns one 10-second 720p clip with sound.
That flatness is the useful engineering property here. Cost estimation collapses to arithmetic: number of clips equals number of successful generations, and budget equals clips multiplied by the per-video rate. No spreadsheet with resolution multipliers, no reconciliation between billed seconds and rendered seconds, no surprise when a long prompt costs more than a short one.
Three practical habits follow from a per-video unit:
Budget in clips, not in seconds
When a product manager asks for "two minutes of footage," translate immediately: that is twelve 10-second clips, and the unit cost is per clip. Your estimate, your rate limiter, and your internal chargeback all speak the same unit.
Count retries explicitly
Every generation that returns a video is billed, including the one you discard because the camera move was wrong. Prompt iteration is a real cost line. Prototype on a small number of clips, lock the prompt, then batch.
Use the price panel as the source of truth
Rates change; this guide does not track them. Read the live number from the pricing panel before you commit a budget, and keep the per-video unit key in your billing code rather than a hardcoded dollar figure scraped from a blog post.
Because both routes share the same billing unit, switching from text-to-video to image-to-video does not change your cost model — which makes it cheap to A/B the two routes on the same brief.
Prompting for Camera, Action, and Sound
The documented prompt guidance names four facets: subject, action, camera movement, sound. On the image route, subject is largely supplied by image_urls, so your prompt carries action, camera, and sound. Write all four into a single prompt string; there is no separate camera field and no separate audio field.
Subject
Name what is on screen and what it looks like: species, material, clothing, era, color. On the image route, keep the subject description short and consistent with your references, then spend your words on motion.
Action
Verbs with a beginning and an end read better than static adjectives. "A hand lifts the lid and steam rises" gives the model a trajectory. Because every clip is 10 seconds, describe an action that plausibly fits in ten seconds — one gesture, one turn, one reveal — rather than a three-act sequence.
Camera movement
State it explicitly: slow push in, static locked-off shot, lateral tracking, handheld follow, gentle orbit. If you say nothing, you have delegated the most visible property of the shot to chance. On the image route, camera language is one of the three things the prompt is documented to describe, so do not skip it.
Sound
Built-in sound is part of every output, so describe it: ambient room tone, wind, footsteps, a specific musical mood, the absence of music. A prompt that specifies sound consistently will beat one that leaves audio to inference, especially across a series of clips that must feel like one piece.
Structuring within 5000 characters
Order the prompt subject → action → camera → sound, and use short sentences. Keep a reusable template per project so a series stays coherent. Iterate by changing one facet at a time, so you can attribute a bad result to the right clause rather than rewriting everything at once.
Common Integration Pitfalls and How to Avoid Them
Most integration bugs around this family come from importing expectations from other video models rather than reading the schema. The recurring ones:
Assuming image_urls produce a pixel-exact first frame
They do not. References guide the video; they are not a guaranteed opening frame. If your product promises "your image, animated," soften the copy, and if you truly need frame-one control, plan an editorial step rather than relying on the model. The upside of the looser contract is creative latitude — your references can be mood, palette, and composition cues rather than a rigid plate.
Sending auto to text-to-video
auto is a valid aspect_ratio only on the image-to-video route. On text-to-video the enum is the seven fixed ratios and the default is 16:9. A shared request builder that defaults to auto for both routes will fail on the text route. Branch the default: 16:9 for text, auto for image.
Mismatched aspect ratio expectations
On the image route, auto follows your references. If you pass nine references with wildly different shapes and then demand a specific ratio, expect the model to reconcile under constraint — sometimes gracefully, sometimes not. The safer pattern: keep reference batches shape-consistent, and only override aspect_ratio when your delivery target genuinely requires it. Once a fixed ratio is chosen, note that the references may be recomposed to fit.
Forgetting the ten-second ceiling in content planning
Scripts written as single 30-second beats will feel truncated. Write in ten-second units from the start, and decide early whether you will stitch clips.
Unvalidated inputs
Enforce prompt length between 1 and 5000 characters, enforce 1–9 image URLs, and confirm each URL is http or https before dispatch. Cheap client-side checks save billed retries.
Putting It Together: A Minimal Integration Checklist
Run this list before your first production request, and again whenever you change a prompt template.
- Pick the route. No existing art →
muse-video-1.0-text-to-video. Approved imagery or a fixed visual style →muse-video-1.0-image-to-video. - Confirm required parameters. Both routes need
prompt(1–5000 chars). The image route also needsimage_urlswith 1–9 http/https URLs. - Set aspect ratio deliberately. Text route: one of
16:9,9:16,1:1,4:3,3:4,3:2,2:3, defaulting to16:9. Image route:auto(default, follows references) or one of the same seven. - Write one prompt covering subject, action, camera movement, and sound — in that order, in short sentences.
- Estimate cost in clips. One generation equals one 10-second video, billed per video at the published rate; multiply by your expected retry count.
- Plan for a fixed artifact. 10 seconds, 720p, with sound, on every successful call. Build thumbnails, captions, and playback around that constant.
- Validate before dispatch and log the route, prompt hash, and reference URLs for every job so bad outputs are traceable to inputs.
The surface area is small, the output is constant, and the two routes differ in a single, legible way — which makes this family a fast one to ship against. Start with the text-to-video route for a first smoke test, then move to image-to-video once you have references worth guiding.
Похожие статьи
Все статьи →-
Guide
gpt-6.2-sol Getting Started Guide: Capability Boundaries, API Integration, and Token Pricing Essentials
Integration decisions and practical essentials for engineering teams
-
Guide
gpt-6.2-astra Getting Started Guide: Text Model Capability Boundaries, Integration Essentials, and Cost Control
-
Guide
Tripo H3.1 Getting Started Guide: Choosing Among Three Input Modes, Parameters, and Pricing
A practical guide for developers on 3D generation integration decisions and cost control