Seedance 2.5 API: Reference-to-Video with Real People, Limits, and Pricing

Seedance 2.5's reference-to-video endpoint now takes real people as references. Here's the full input spec, the price by resolution, and when to use it over image-to-video.

Seedance 2.5 has three endpoints on Yotta AI Gateway: text-to-video, image-to-video, and reference-to-video. The third one is the interesting one. Instead of a single starting frame, you hand it a set of reference assets, images, video clips, and audio, and describe in the prompt how they should come together. Seedance calls these omni references, and 2.5 accepts up to 50 of them in one request.

Two things changed recently that make this worth a post of its own. Seedance 2.5 on the Gateway now accepts real people as reference subjects, so a photo or clip of an actual presenter, founder, or customer can anchor a generated video. And some earlier restrictions on reference inputs have been relaxed. This post covers what the endpoint accepts today, what it costs, and how it compares to Seedance 2.0 and to the simpler image-to-video route.

The short version

• Real people as references: now accepted on Seedance 2.5 reference-to-video on Yotta AI Gateway.
• Inputs per request: up to 30 images, 10 video clips, and 10 audio clips, 50 assets total.
• Output: 4 to 30 seconds, 480p to 1080p, mp4 or mov, audio generated by default.
• Price: from $0.08 per second at 480p with a video reference to $0.59 per second at 1080p with none.
• Against Seedance 2.0: three times the reference capacity and double the clip length, at a higher per-second rate; 2.0 still has 4K.

What the reference-to-video endpoint accepts

Everything below is from the Seedance 2.5 reference-to-video model page on the Yotta console.

At least one reference asset is required. The prompt does the directing: which image is the first frame, who the person in Image 2 is, what the audio clip is for, how the camera moves. Seedance 2.5 also supports multi-round extension and timestamp-based edits on the result, which is how the 30-second ceiling becomes longer pieces.

What changed: real people as references

Until recently, the practical limit on reference-to-video was that your reference subjects had to be products, scenes, characters, or styles. Seedance 2.5 on the Gateway now accepts real people as reference subjects, and some of the earlier restrictions on reference inputs have been relaxed.

What that opens up is the most common commercial use of the format. A product shot, a short clip of the actual person who presents it, and a voice sample, combined into a spot where the real presenter appears with consistent features across the clip. The same applies to founder updates, customer stories, and talking-head explainers that want a specific face rather than a generated one.

Two practical notes. The person's likeness should be one you have permission to use, which is on you as the caller, and Seedance's content rules still apply to the prompt and the output. And give the model clean references: a front-facing image with the face clearly visible, and a short clip with normal lighting, work far better than a crowd shot.

Seedance 2.5 reference-to-video pricing

Per second of generated video, from the model page. "With video input" means at least one video clip is among the references.

What that means for a few common jobs:

Two things to notice. The video-reference rate is the cheaper one, so if you have a clip of the subject, use it; it costs less and the result is more consistent. And resolution is the expensive dial: 1080p is four to five times the 480p rate, so draft at 480p or 720p and render the final at 1080p.

Seedance 2.5 vs Seedance 2.0 for reference-to-video

Both are on the Gateway. The difference is capacity and price.

Use 2.5 when the job needs a lot of references, clips longer than 15 seconds, or the best character consistency. Use 2.0 when you need 4K output or you're generating a high volume of short clips and the per-second rate matters more than the ceiling. The Seedance 2.5 vs 2.0 post goes through the quality differences in more detail.

Reference-to-video or image-to-video?

Seedance 2.5 image-to-video takes one starting image and a prompt, and costs $0.12, $0.25, or $0.59 per second at 480p, 720p, and 1080p, the same as reference-to-video without a video input. So the choice isn't about price. It's about control.

Image-to-video is the right call when you have one frame and want it to move: a product render, a still from a shoot, a storyboard panel. Reference-to-video is the right call when the video has to feature specific things that don't all fit in one frame: a person, a product, a setting, and a voice. If you're deciding by the number of things that must look right, one thing means image-to-video and more than one means reference-to-video.

Making the request

The Gateway serves video models through an async API: submit a job, get a request ID, poll for the result. The base URL for video and image models is https://gateway.yottalabs.ai/api/maas and the key goes in the X-API-KEY header.

A reference-to-video submission is a POST to /reference-to-video/generations with the model slug seedance-2.5-r2v and a parameters object. Every reference goes in reference_items as a URL plus a type of image, video, or audio. A minimal payload looks like this:

{
  "model": "seedance-2.5-r2v",
  "parameters": {
    "prompt": "Image 1 is the presenter. She introduces the product in Image 2 against a bright studio background, camera slowly pushing in. Use Audio 1 for her voice.",
    "reference_items": [
      {"url": "https://example.com/presenter.jpg", "type": "image"},
      {"url": "https://example.com/product.png", "type": "image"},
      {"url": "https://example.com/voice.mp3", "type": "audio"}
    ],
    "resolution": "720p",
    "duration": 12,
    "aspect_ratio": "16:9"
  }
}

The response carries a request_id. Poll GET /reference-to-video/generations/{request_id} until the status is completed, and read the video from output_url in the result. The console's sample prompt also shows how to direct a long clip: break the prompt into timed sections like [0-10s] and [10-20s] and describe each one. The Seedance API how-to has the full client with polling and error handling, and the console model page has the same code ready to copy.

One more step that's worth knowing: the Gateway also carries Topaz Labs Starlight for video upscaling, so a 720p draft from Seedance can be upscaled on the same API key. The Starlight Precise 2.6 vs Fast 3 post covers that step.

Frequently asked questions

Can Seedance 2.5 use a real person as a reference? Yes. On Yotta AI Gateway, Seedance 2.5 reference-to-video now accepts real people as reference subjects. Use likenesses you have permission to use.

How many reference assets does Seedance 2.5 accept? Up to 50 per request: 30 images, 10 video clips, and 10 audio clips. Video and audio references are capped at 30 seconds combined per type.

How long a video can Seedance 2.5 generate? 4 to 30 seconds in one request, with multi-round extension for longer pieces.

How much does the Seedance 2.5 API cost? Reference-to-video runs $0.08 to $0.59 per second depending on resolution and whether a video clip is among the references. Image-to-video is $0.12 to $0.59 per second. A 15-second 720p clip with a video reference is $2.25.

Is Seedance 2.5 cheaper than Seedance 2.0? No. Seedance 2.0 is cheaper per second ($0.10 against $0.15 at 720p with a video reference) and offers 4K. Seedance 2.5 takes three times the references and makes clips twice as long.

What is the Seedance 2.5 API endpoint? POST /reference-to-video/generations on https://gateway.yottalabs.ai/api/maas with model seedance-2.5-r2v, then poll the request ID.

Does Seedance 2.5 generate audio? Yes, by default. Set generate_audio to false to get silent video, or supply an audio reference to steer the sound.

Can I get 4K from Seedance 2.5? Not on reference-to-video; the ceiling is 1080p. Seedance 2.0 offers 4K, and Topaz Labs Starlight on the Gateway can upscale a 1080p result.

Bottom line

Seedance 2.5 reference-to-video on Yotta AI Gateway now takes real people as references, which turns it from a product-and-scene tool into one that can carry a presenter through a spot. The endpoint accepts up to 50 assets and makes 30-second clips at up to 1080p, priced from $0.08 a second when you include a video reference.

Draft at 720p with a video clip of the subject, render the final at 1080p, and upscale on the same key if you need more. For a single moving frame, image-to-video at the same price is the simpler route. Both are one API key away.