muse-videoMuse Video is the video model from Meta Superintelligence Labs. On APIMODELS it is one model id, muse-video, on the unified video endpoint: send a prompt for text-to-video, add one first-frame image and the clip starts from that picture, or add up to four reference images and the model keeps the same person or subject without starting from them. The request stays small. prompt is required (up to 4,000 characters, any language). duration sets the length in whole seconds from 4 to 30 (the OpenAI-style seconds field works too); leave it out and you get 5 seconds. aspect_ratio picks 16:9 landscape (the default, 1280×720) or 9:16 portrait (720×1280); size 1280x720 or 720x1280 does the same. Output is always 720p: there is no 1080p or square option, and asking for one returns a 400 before any task is created. For image-to-video, pass the first frame as image_url (first_frame_url, image and input_reference also work), ideally in the same orientation as the clip; for reference images, pass up to four in images (image_urls and reference_image_urls also work). Each image is a public URL or a base64 data URI in PNG, JPEG or WebP. Use one or the other: a first frame together with reference images, more than four reference images, a last frame, reference video or audio, negative_prompt, seed and turning sound off are rejected with a 400. Clip lengths come in steps. Requests of 5, 6, 8, 10, 15, 20, 25 or 30 seconds come back exactly that long; any other whole number rounds up to the next step (4 → 5 s, 7 → 8 s, 11 → 15 s), so a clip is never shorter than what you paid for. The video is an H.264 MP4 at 24 fps, and Meta describes Muse Video as generating sound together with the picture. In our reference-image test, four pictures of one woman produced a 5-second beach clip in which she kept the same face and hair; the model did not start from any of the pictures. It runs as an async task: create it, then poll or take a callback. Clips up to 10 seconds usually take one to two minutes, 15 to 30 seconds three to seven. Not every request succeeds: the model can refuse content, a generation can hit a temporary error, and a clip with the wrong orientation or length is discarded; a failed task is not charged. Pricing is $0.037 per requested second: 5 s $0.185, 10 s $0.37, 30 s $1.11.
The API also takes any other whole number from 4 to 30; it rounds up to the next length here (7 s comes out as 8 s) and is billed on the seconds you request.
Generated video will appear here
Provide URLs and click Generate
One model id: a prompt alone, one first frame to start the clip from your picture, or up to four reference images to keep the same person across clips
Whole seconds; 5, 6, 8, 10, 15, 20, 25 and 30 s come out exact, other lengths round up, so a clip is never shorter than what you pay for
1280×720 (16:9, default) or 720×1280 (9:16), H.264 MP4 at 24 fps; no 1080p, no square or other ratios
Meta describes Muse Video as generating audio together with the video
Not every request succeeds, and a failed task is not charged; unsupported lengths, sizes or inputs are rejected with a 400 before a task is created
Billed on the requested seconds: 5 s $0.185, 10 s $0.37, 30 s $1.11
Muse Video is Meta's video model. Send a prompt alone, add one first-frame image to start the clip from that picture, or add up to four reference images to keep the same person or subject, and get a 720p MP4 of 4 to 30 seconds in 16:9 landscape or 9:16 portrait.
On APIMODELS it is one model id, muse-video, on POST /v1/video/generations, billed at $0.037 per requested second. Check the limits below before you build on it: 720p only, landscape or portrait only, one first frame or up to four reference images but not both, and not every request succeeds.
muse-videoText-to-video, no first frame, 9:16, 10 seconds. One generation, shown as it came back.
One request with the prompt below, duration 10 and aspect_ratio 9:16. Muse Video returned a 10.0-second 720×1280 clip with a music track in about 75 seconds, and it played the action list in order in one continuous take: eyes closed on the pillow, eyes opening to the lens, a shy smile, the blanket pulled up over half the face, a V sign close to the lens, a hand over the mouth while she laughs into the pillow, and the phone tipping toward the ceiling.
This is the kind of clip Muse Video handles well: one person, a handheld vertical selfie frame, a short chain of small gestures, and a mood carried by warm low light and music. Another run of the same prompt will look different. Text-to-video cannot keep the same face from clip to clip, so for a recurring character, pass up to four reference images of them (images); the model keeps the face without starting from the picture.
主体：人物是明月（我的固定角色），穿奶白低领细肩吊带；头发被枕头压乱，多缕发丝横过脸部。从闭眼侧躺开始，再睁眼互动。全程同一张脸、同一套衣服。
场景：夜间卧室床上，白被子白枕头，暖色床头灯。她自己拿手机前置超近距自拍，头到上胸几乎占满竖屏，腰腹和下半身不入画。手腕轻轻晃，构图略歪；一镜到底，不切镜。
音乐：轻甜 bedroom pop 器乐，约 90 BPM。无台词、无说话口型。配乐里不要人声和笑声。
动作：侧脸贴枕闭眼浅笑 → 慢慢睁眼锁定镜头 → 下巴抵肩害羞抿嘴笑 → 拉起白被子挡住半张脸再拉开 → 空着手从上胸前伸向镜头，V 手势几乎贴住镜头再收成夹捏/小爱心 → 握拳挡嘴害羞偷笑，把脸埋进枕头 → 手机一歪甩向天花板。实时速度。
画质：故意拍糊，手持动态模糊和局部失焦，暖色低照度，10 秒竖屏。What Muse Video accepts and what it does not. Unsupported resolutions, ratios, lengths and inputs are rejected with a 400 before a task is created, so nothing is charged.
| Supported | Not supported | |
|---|---|---|
| Resolution | 720p only (resolution: 720p, or leave it out) | 1080p, 2K, 4K, 480p |
| Aspect ratio | 16:9 landscape, 1280×720 (default); 9:16 portrait, 720×1280 | Square 1:1, 4:3, 3:4, 21:9 and every other ratio |
| Length | 4 to 30 whole seconds; 5, 6, 8, 10, 15, 20, 25 and 30 s come out exact, other lengths round up to the next of those | Under 4 s, over 30 s, fractions such as 8.5 |
| Images | Either 1 first frame, or up to 4 reference images that keep the same person or subject (not both); PNG, JPEG or WebP, up to 20 MB each, public URL or base64 | A first frame together with reference images, more than 4 reference images, a last frame |
| Video and audio input | None: images are the only media input | Reference video, video editing or extension, driving audio or lip sync to your own audio |
| Sound | Generated by the model together with the picture | A switch to turn sound off (generate_audio: false is rejected); sound is not guaranteed on every clip |
| Other controls | A prompt of up to 4,000 characters in any language; describe camera moves in the prompt | negative_prompt, seed (results are not repeatable), camera-control parameters |
Like every video model, Muse Video does not succeed 100% of the time. Failed tasks are not charged.
A task can still fail after it is accepted: the model can refuse the content, a generation can hit a temporary error, or a finished clip that does not match the requested orientation or length is discarded. A failed task ends in the failed state with the reason in failMsg, and the charge is returned.
Both of our launch test requests returned a clip, which is far too few to promise a success rate, so we do not publish one yet. Plan your pipeline for the occasional failure: retry temporary errors, rewrite refused prompts, and do not assume a clip will arrive on the first attempt.
The model can refuse a prompt or an input image. Retrying the same request rarely helps; rewrite the prompt or change the image.
A generation can fail for reasons unrelated to your request. Retrying the same request a little later usually works.
A finished clip with the wrong orientation or a shorter length than requested is discarded instead of delivered. Retry the request.
Meta describes sound generated with the picture, but one of our two test clips came back silent. If you need audio, check each MP4 for an audio track.
$0.037 per requested second. A first frame or reference images cost nothing extra; failed tasks are free.
| You request | Clip you get | Price |
|---|---|---|
| 4 s | 5 s | $0.148 |
| 5 s (default) | 5 s | $0.185 |
| 6 s | 6 s | $0.222 |
| 7 s | 8 s | $0.259 |
| 8 s | 8 s | $0.296 |
| 10 s | 10 s | $0.37 |
| 15 s | 15 s | $0.555 |
| 20 s | 20 s | $0.74 |
| 25 s | 25 s | $0.925 |
| 30 s (maximum) | 30 s | $1.11 |
Clips up to 10 seconds usually take one to two minutes and 15 to 30 seconds three to seven minutes; our 4-second test requests took about a minute and a half. Clips over 10 seconds switch shots around the 10 and 20-second marks.
One person filmed close up on a front camera, a chain of small gestures, warm low light and music, like the example above.
Pass up to four photos of the same person in images and describe a new scene. In one test, four photos of a young woman and a sunset-beach prompt gave a 5.0-second 9:16 clip with sound in 88 seconds: the same face and hair, a similar outfit, even a V sign and a laugh behind her hand taken from the photos. It opened on the beach, not on a photo. One test, not a guarantee.
9:16 clips for short-video feeds or 16:9 clips for pages and ads, 5 to 30 seconds at 720p.
A product shot, an illustration or a photo as the first frame, with the motion described in the prompt.
Muse Video only renders 720p in 16:9 or 9:16. Choose another video model on APIMODELS if you need higher resolution or other shapes.
Images go in as a first frame or as reference images, and that is all: no last frame, no reference video, no lip sync to your own audio.
Some requests fail and some clips come back without sound. If a missed clip breaks your product, add retries and checks before you rely on it.
Muse Video is a Video Generation API provided by Meta. Muse Video is the video model from Meta Superintelligence Labs. On APIMODELS it is one model id, muse-video, on the unified video endpoint: send a prompt for text-to-video, add one first-frame image and the clip starts from that picture, or add up to four reference images and the model keeps the same person or subject without starting from them. The request stays small. prompt is required (up to 4,000 characters, any language). duration sets the length in whole seconds from 4 to 30 (the OpenAI-style seconds field works too); leave it out and you get 5 seconds. aspect_ratio picks 16:9 landscape (the default, 1280×720) or 9:16 portrait (720×1280); size 1280x720 or 720x1280 does the same. Output is always 720p: there is no 1080p or square option, and asking for one returns a 400 before any task is created. For image-to-video, pass the first frame as image_url (first_frame_url, image and input_reference also work), ideally in the same orientation as the clip; for reference images, pass up to four in images (image_urls and reference_image_urls also work). Each image is a public URL or a base64 data URI in PNG, JPEG or WebP. Use one or the other: a first frame together with reference images, more than four reference images, a last frame, reference video or audio, negative_prompt, seed and turning sound off are rejected with a 400. Clip lengths come in steps. Requests of 5, 6, 8, 10, 15, 20, 25 or 30 seconds come back exactly that long; any other whole number rounds up to the next step (4 → 5 s, 7 → 8 s, 11 → 15 s), so a clip is never shorter than what you paid for. The video is an H.264 MP4 at 24 fps, and Meta describes Muse Video as generating sound together with the picture. In our reference-image test, four pictures of one woman produced a 5-second beach clip in which she kept the same face and hair; the model did not start from any of the pictures. It runs as an async task: create it, then poll or take a callback. Clips up to 10 seconds usually take one to two minutes, 15 to 30 seconds three to seven. Not every request succeeds: the model can refuse content, a generation can hit a temporary error, and a clip with the wrong orientation or length is discarded; a failed task is not charged. Pricing is $0.037 per requested second: 5 s $0.185, 10 s $0.37, 30 s $1.11. Through APIMODELS platform, you can access this model via a unified API with transparent pay-as-you-go pricing. Current pricing: per second: $0.037, 5s (default): $0.185, 8s: $0.296, 10s: $0.37, 15s: $0.555, 20s: $0.74, 30s (maximum): $1.11.
Quickly generate brand promotion videos for ad campaigns and social media marketing.
Create compelling short-form video content for platforms like TikTok, Instagram, and YouTube.
Generate product feature demonstrations and tutorials to improve user conversion.
Produce course explanations, knowledge explainers, and training videos at an affordable price.
Muse Video is available through APIMODELS at: per second: $0.037, 5s (default): $0.185, 8s: $0.296, 10s: $0.37, 15s: $0.555, 20s: $0.74, 30s (maximum): $1.11. Billing is pay-as-you-go — you only pay for what you generate.
Sign up at APIMODELS, get your API key, and call our unified API endpoint. We provide detailed API documentation with code examples in cURL, Python, and Node.js.
APIMODELS offers the same Muse Video model through our aggregation platform. We provide a unified API interface so you do not need separate accounts for each provider - one API key to access all models.
Muse Video is the video model from Meta Superintelligence Labs. It turns a text prompt into a 720p clip of 4 to 30 seconds in 16:9 landscape or 9:16 portrait; add one first-frame image and the clip starts from that picture, or add up to four reference images to keep the same person or subject. Meta describes it as generating sound together with the picture. On apimodels.app, Meta's Muse Video model is available as one model id, muse-video, on POST /v1/video/generations. Failed generations are not charged, and pricing is $0.037 per requested second.
All of Muse Video's limits: 720p only (no 1080p, 2K, 4K or 480p); 16:9 landscape at 1280×720 or 9:16 portrait at 720×1280 only (no square, 4:3, 21:9 or any other ratio); 4 to 30 whole seconds, exact at 5, 6, 8, 10, 15, 20, 25 and 30 s, other lengths rounded up, no fractions; images: either one first frame or up to four reference images that keep the same person or subject, not both, each PNG, JPEG or WebP up to 20 MB (no last frame and no more than four reference images); no video input, no audio input, no lip sync to your own audio and no video editing or extension; Meta describes the model as generating sound, but sound is not guaranteed on every clip and cannot be turned off; no negative_prompt, no seed and no camera-control parameters (describe camera moves in the prompt); prompts up to 4,000 characters. Unsupported inputs, such as another resolution or ratio, a first frame together with reference images, more than four reference images, a last frame, video or audio input, negative_prompt, seed or generate_audio: false, are rejected with a 400 before a task is created. Even within these limits, not every request succeeds: the model can refuse content or hit a temporary error, and a failed task is not charged.
Yes. One model id, muse-video, has three input modes, picked from what you send. A prompt alone is text-to-video. One first-frame image makes the clip start from that picture: pass it as image_url (first_frame_url, image and input_reference also work). One to four reference images keep the same person, subject or style without the clip starting from any of them, which suits a recurring character: pass them as an array in images (image_urls and reference_image_urls also work). Any array counts as reference images, so image_urls with one item is one reference image, not a first frame. Send a first frame or reference images, not both: both together, or more than four reference images, return a 400 before a task is created. Every image is a public URL or a base64 data URI in PNG, JPEG or WebP, up to 20 MB. A first frame works best in the same orientation as the clip, landscape for 16:9 and portrait for 9:16. A last frame (last_frame_url), reference video and audio input are not supported. All three modes cost the same.
Reference images tell the model who or what should appear; they do not become frames of the video. Pass one to four images of the same person or subject in images, describe the new scene in the prompt and point back to them, for example "the same young woman from the reference images". We tested this once, on 30 October 2026: four photos of one young woman (lying on a pillow, smiling, making a V sign, laughing behind her hand) with the prompt "The same young woman from the reference images walks barefoot along a beach at sunset, smiling back at the camera, handheld vertical shot", 4 seconds at 9:16. After 88 seconds it returned a 5.0-second 720×1280 clip with an audio track. The clip did not start from any of the reference images; it opened on the beach. The woman kept the same face and hair and a similar outfit (the cream camisole in the references became a white slip dress), the model reused gestures from the references (the V sign and laughing behind a hand), and it cut between several shots within the 5 seconds: close-ups, a wide shot and a back view. That is one test, not a guarantee of likeness or of how the clip is cut. If the clip must open on an exact picture, use a first frame instead. Reference images cost nothing extra: the 4-second request was billed exactly like a 4-second text-to-video request.
From 4 to 30 seconds, set in whole seconds with duration (the OpenAI-style seconds field also works); leave it out and you get 5 seconds. Output lengths come in steps: requests of 5, 6, 8, 10, 15, 20, 25 or 30 seconds come back exactly that long, and any other whole number rounds up to the next step: 4 → 5 s, 7 → 8 s, 9 → 10 s, 11–14 → 15 s, 16–19 → 20 s, 21–24 → 25 s and 26–29 → 30 s. A clip is never shorter than what you paid for. If you need an exact length, request one of the steps. Fractions and values outside 4–30 return a 400 before a task is created.
720p only, in two shapes: 16:9 landscape at 1280×720, the default, and 9:16 portrait at 720×1280. Choose with aspect_ratio set to 16:9 or 9:16, or with size 1280x720 or 720x1280; resolution is optional and accepts only 720p. There is no square, other ratio or 1080p option, and asking for one returns a 400 before a task is created instead of quietly delivering a different size. Every clip is an H.264 MP4 at 24 fps.
Meta describes Muse Video as generating sound together with the picture. In our own launch tests on 30 October 2026, two 4-second requests came back as two 5.0-second clips: the landscape one carried an AAC audio track and the portrait one came back silent. So if your pipeline needs sound, check each MP4 for an audio track (ffprobe shows it) and plan a fallback, such as adding music or a voice-over, for clips that arrive without one. Sound cannot be switched off: generate_audio: false returns a 400 ("muse-video cannot turn sound off: the model decides whether a clip has sound"), and generate_audio: true is accepted but does not guarantee sound.
Clips up to 10 seconds usually take one to two minutes, and clips of 15 to 30 seconds usually three to seven minutes; our 4-second test requests finished in about 1.5 minutes. The call is asynchronous: creating a task returns a taskId at once, then you poll GET /v1/video/generations?task_id= every 10 to 15 seconds, or pass callback_url and wait for our POST.
$0.037 per requested second, billed on the duration you send: 4 seconds $0.148, 5 seconds (the default) $0.185, 6 seconds $0.222, 8 seconds $0.296, 10 seconds $0.37, 15 seconds $0.555, 20 seconds $0.74, 25 seconds $0.925 and 30 seconds $1.11. A length between the steps is billed on the seconds you asked for: 7 seconds costs $0.259 and comes back as an 8-second clip. A first frame or reference images cost nothing extra, so all three input modes are priced the same, and the aspect ratio does not change the price. Failed generations are not charged.
Not every Muse Video request succeeds. A generation can fail for three reasons: the model refuses the content, the generation hits a temporary error, or the finished clip has the wrong orientation or is shorter than requested and is discarded. A failed task ends in the failed state with the reason in failMsg, and you are not charged. We do not publish a success rate yet: both of our launch test requests returned a clip, but two tests are far too few to promise one. Retry temporary errors, rewrite a refused prompt, and do not build a pipeline that assumes the first attempt always delivers. Requests Muse Video cannot serve never become tasks: a missing prompt or one over 4,000 characters, a duration that is not a whole number from 4 to 30, a resolution other than 720p, an aspect ratio other than 16:9 or 9:16, a first frame together with reference images, more than four reference images, a last frame, video or audio input, negative_prompt, seed or generate_audio: false all return a 400 with a plain error message before a task is created. The images themselves are checked at run time: one that is not PNG, JPEG or WebP, or is larger than 20 MB, fails the task with the reason, such as "first-frame image must be PNG, JPEG or WebP" or "reference image 2 must be PNG, JPEG or WebP", also without charge.
Say what the subject is, how it moves and how the camera moves, for example "an orange cat stretches on a windowsill, sunlight pours in, the camera slowly pushes in". Prompts can run to 4,000 characters in any language. Clips longer than 10 seconds switch shots around the 10-second and 20-second marks, so write a long prompt segment by segment: "first 10 seconds …, then …". For image-to-video, use a first frame in the same orientation as the clip and let the prompt describe what happens next. With reference images, call the subject "the same person from the reference images" and describe the new scene, since the clip will not start from any of them.
Create an account at apimodels.app and generate an API key in the console; one key and one USD balance cover every model. Then send: curl -X POST https://api.apimodels.app/v1/video/generations -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" -d '{"model": "muse-video", "prompt": "an orange cat stretches on a windowsill, sunlight pours in, the camera slowly pushes in", "duration": 5, "aspect_ratio": "16:9"}'. Add "image_url" to start from your own first frame, or "images" with up to four reference images to keep the same person or subject. The response holds a taskId: poll GET /v1/video/generations?task_id= every 10 to 15 seconds until the state is completed, or pass callback_url, then download the MP4 from resultUrls within 30 days.
Not entirely. Eligible new accounts get $0.10 of trial credit, but the shortest Muse Video request, 4 seconds, costs $0.148, so a first test needs a top-up; top-ups start at $10. Billing is pay-as-you-go with no subscription. The playground on this page shows the price before you click Generate, and failed generations are not charged.
On APIMODELS, Muse Video runs alongside 60+ models on one API key and one balance, so choosing is about fit, not lock-in. It supports Text to Video, Image to Video (First Frame), Up to 4 Reference Images, 720p · 16:9 / 9:16, 4–30s · Exact at 5/6/8/10/15/20/25/30s, Sound with the Picture, Failed Generations Not Charged, $0.037/s, and you can weigh it on price and capability against other Video Generation models, then switch by changing a single model-name string — no new account or integration. Browse every Video Generation option with live pricing at apimodels.app/models.
Muse Video supports: Text to Video, Image to Video (First Frame), Up to 4 Reference Images, 720p · 16:9 / 9:16, 4–30s · Exact at 5/6/8/10/15/20/25/30s, Sound with the Picture, Failed Generations Not Charged, $0.037/s. See the APIMODELS docs for full parameters and call examples.
Yes. APIMODELS exposes Muse Video through a single unified API and one key — no separate provider accounts, and no need to handle each provider's regional network access yourself.
We support Stripe (Visa, Mastercard, and other international cards) and Alipay. Credits are available instantly after payment.