# Seedance AI 2.5: Specs, Cost, Filters and When to Use It for Product Videos

> A practical guide to Seedance 2.5 for founders and devrel teams: how the talking avatar works, what it costs per second, what its content filter blocks, and when Kling or Veo fits better.

**Key takeaways**

- Seedance 2.5 renders a full 30-second clip in one generation and natively tops out at 1080p, with true 1440p only via a separate enterprise-only 4K model.
- On BytePlus ModelArk it costs $0.1028 per second at 480p, $0.2311 at 720p and $0.5686 at 1080p, pricier per second than Kling 3.0 or Veo 3.1.
- Seedance's face filter refuses any generation using a real person's photo as a reference, so presenters and characters must be described in words instead.

Published Oct 23, 2026 · uscale · https://uscale.ai/blog/seedance-2-5-product-videos

## Thirty seconds, one take

Seedance AI's newest model, Seedance 2.5, renders a full 30-second clip in a single generation: no stitching shorter segments together, no extension passes. Its predecessor managed roughly 15 seconds natively. For a founder planning a product launch video, that doubling matters more than most spec-sheet lines. Thirty seconds is a whole teaser, a presenter's pitch to camera, or a two-line exchange between a founder and a customer. On paper, one prompt can now cover what used to take a timeline.

[Video: Founder film for RoiMaxi](https://uscale.ai/examples/roimaxi-founder-film.mp4) — A founder avatar film made in uscale — one photo, a voice track, lines to camera.
Then the details arrive. That single take is billed by the second. On BytePlus ModelArk, ByteDance's official platform, the 1080p rate is $0.5686 per second without video input. And the same model that holds a scene together for half a minute refuses to start at all if the reference photo is a real person's face. So the open question for a launch video is not whether Seedance 2.5 can make thirty seconds. It is which thirty seconds are worth handing it.

Answering that starts with a plainer question: what Seedance is, and where you actually meet it.

## What is Seedance AI, and who owns it?

Seedance is ByteDance's AI video model. Its current version, Seedance 2.5, has a stable release date of [July 31, 2026](https://en.wikipedia.org/wiki/Seedance). It reached people in a set order. It launched first as the default video model inside ByteDance's own Jimeng and Dreamina apps, and [developer API access on Volcano Engine and BytePlus followed](https://www.miracamp.com/learn/content-creation/seedance-2-5-bytedance).

That order matters more than the date. The same model reaches you through two doors, and each one fits a different job. In an app, someone else has already made the choices around the model. Through the API, you write the request yourself, and you take on the model's rules, limits and billing directly.

| Where you use Seedance 2.5 | What it is | Who it suits |
|---|---|---|
| Jimeng and Dreamina | ByteDance's own apps, where Seedance 2.5 is the default video model | A founder who wants a clip without writing code |
| Volcano Engine and BytePlus | ByteDance's developer platforms, where API access came after the app release | A builder calling the model from their own tool or pipeline |

Here is a rule worth keeping for any model you look at. Before you trust a comparison, find out which door it came through. A result from an app and a result from the API can come from the same model with different defaults around it. Most of the disagreement you read about a new model comes from two people using it in two different places.

For a product launch video, the door you pick also decides how you pay. That is the first practical question: what one second of Seedance 2.5 costs.

## Is Seedance AI free, and what does it cost per second?

No. On the official API, Seedance 2.5 is billed by the second, and no free rate is listed with the paid ones. Inside ByteDance's own Jimeng and Dreamina apps, it ships as the default video model. What those apps charge is set by the apps, not by the API price list.

To plan a budget, use the rate on BytePlus ModelArk, ByteDance's official platform. ModelArk added an official 1080p price on a pricing page last updated October 8, 2026, so all three tiers below come from the same source.

![Seedance 2.5 price per second on BytePlus ModelArk: 480p $0.1028/s, 720p $0.2311/s, 1080p $0.5686/s](https://uscale.ai/blog-media/seedance-2-5-product-videos/chart-seedance-2-5-price-per-second-on-byteplu-1200x400.svg "Source: ArgoLink / Seedance 2.5 API Pricing: 6 Providers Compared, Oct 6, 2026")
These rates cover requests without video input. If you give the model a reference video clip, these figures don't apply, so check that price before you plan a scene built on existing footage.

### What that means for a 10–30 second clip

You control two things: length and resolution. At the same tier, a 30-second shot costs three times as much as a 10-second one. Each step up in resolution more than doubles the per-second rate, so resolution moves the bill more than length does.

Most founders miss the cost of discarded takes. You pay for every second rendered, not only the seconds you keep. A shot that takes four tries to get right costs four clips. **Find the shot at 480p, then render it once at 1080p.** The [Seedance 2.5 draft-mode guide](https://uscale.ai/blog/seedance-2-5-draft-mode-product-videos) covers that workflow, and the [video cost calculator](https://uscale.ai/tools/video-cost-calculator) shows how takes and length add up.

The presenter clip in the next section runs on a separate model with its own price band. Compare it against these rates before you choose.

## How does the Seedance 2.5 talking avatar turn a photo into a presenter?

Seedance 2.5 has a separate [Talking Avatar model](https://wavespeed.ai/collections/avatar-lipsync) for this job. You give it one portrait and an audio recording, and it returns a video of that face speaking the recording at 480p or 720p. It reads up to the first 120 seconds of the audio, so anything past that mark is not animated.

That gives you a presenter shot without a studio: one person facing the camera, saying a line about your product. The model works from one portrait and one track. If you need several people in a scene, use the main Seedance 2.5 model instead.

### The audio is the script

Most people treat the photo as the hard part. It is the easy part. The audio sets the length, the pacing and every word the mouth forms. **A talking avatar is only as good as the line you recorded before you pressed generate.** Settle the words first, because changing one sentence means a new generation.

### What it costs

The Talking Avatar model is [priced at $0.22 to $0.44 per second](https://www.mindstudio.ai/models/seedance-2-5-talking-avatar). The low end sits close to the general model's 720p rate from the previous section. Its output stops at 720p, so plan for it as a 720p tool with its own price band.

### How a founder avatar film uses it

A uscale founder avatar film is built around this kind of shot. One clear photo becomes a presenter who speaks your approved lines to camera, with your product composed around them. Each line runs up to about 30 words, which is 4 to 15 seconds to camera. A film has up to six presenter lines, and every line is approved in the storyboard before anything is filmed. The film below uses our own test photo, not a customer's.

[Video: Founder film for Northpoint Studio](https://uscale.ai/examples/northpoint-founder-film.mp4) — Another founder avatar film made in uscale: the presenter photo is ours, the lines are approved in the storyboard first.
One face and one voice cover a presenter line. A scene with two speakers needs a different brief, and the next section covers it.

## How do you write a Seedance 2.5 skit with dialogue?

In a Seedance 2.5 skit, the audio track keeps the scene in sync, not the text prompt. Seedance 2.5 adds an audio-only reference: a voice, music or sound-effect track that can drive pacing, beat-matching and lip-sync, with native support for more than ten languages. This changes how you brief a two-line exchange. Make the dialogue recording first and let the model fit the picture to it. Don't describe the timing in words and hope the mouths land.

The model can also take far more references than before. Each generation accepts up to 50 inputs, split as shown below. Seedance 2.0 took 9 images and 3 audio or video clips.

![Seedance 2.5's reference budget per generation: Images 30 images, Video clips 10 video clips, Audio clips 10 audio clips](https://uscale.ai/blog-media/seedance-2-5-product-videos/chart-seedance-2-5-s-reference-budget-per-gene-1200x396.svg "Source: Morphic / Seedance 2.5: edit, extend, and native 1080p AI video, Jan 1, 2026")
### What to hand the model

A short dialogue scene needs only a small part of that budget:

- **One reference image per speaker**, so each character stays the same person from start to end. Use generated or illustrated characters. Seedance's face filter rejects any generation that uses a real person's photo as a reference.
- **One audio track that holds both lines**, so timing and lip-sync come from a single source.
- **A text prompt** that names the setting, the camera and which character speaks first.

Put both lines in one track. Don't split them into two clips. One track means one clock: the pause between the lines is already in the recording. With two clips, the model has to decide how long the gap is. Record the conversation first. The audio is the timeline, and the picture follows it.

Here is an example. Say you're briefing a skit for a launch. One character asks, "How long did that export take?" and the other answers, "It's already done." You give the model two character images, one recording of both lines with a beat of silence after the answer, and a prompt that sets the scene in an office. The single take renders without stitching, so the question, the answer and the reaction can all sit in one shot, with no join between segments to plan around.

Once the scene is in sync, the next question is how sharp it looks on screen. That is where the resolution claims need the closest reading.

## What resolution does Seedance 2.5 actually deliver — and is 1440p real?

Seedance 2.5 renders natively at 1080p: a full 1920x1080 frame drawn directly by the model, up from 720p at launch. That is the highest resolution the generally released model makes. Any "1440p" Seedance 2.5 output you see quoted is that same 1080p render, made bigger.

[Atlas Cloud's own model page](https://www.atlascloud.ai/models/bytedance/seedance-2.5/image-to-video) says so. Its "1440p-sr" and "1440p-esr" options take the nearest native 1080p source and scale it to a 1440-pixel short edge. The file gets more pixels, but the model did not draw more detail. An upscaler can smooth edges. It cannot recover button lettering that the 1080p render left soft.

| Tier | What it is | Native render? | Who can use it |
|---|---|---|---|
| 1080p | Full 1920x1080 frame rendered directly | Yes | The released Seedance 2.5 model |
| 1440p-sr / 1440p-esr | Nearest 1080p source scaled to a 1440-pixel short edge | No, upscaled | Atlas Cloud option |
| Seedance 2.5 4K | Separate model for film and TV production, released October 23, 2026 | Separate model | Selected enterprise customers, trial only |

The only real step above 1080p is a different model. On October 23, 2026, ByteDance's Volcano Engine released [Seedance 2.5 4K](https://technode.com/2026/10/23/bytedance-launches-seedance-2-5-4k-video-model-for-film-production/), aimed at film and television production. Only selected enterprise customers can use it, and only as a trial. If you are comparing resolutions for a launch video you need soon, leave it out.

### What to expect in practice

Judge a resolution by where its pixels were made, not by the number on the export. For a product video, plan around the native 1080p render. If a shot's UI text, logo or presenter face looks soft at 1080p, a 1440p upscale will show the same softness, only larger. Fix it at the source with tighter framing or a cleaner reference. The upscale option will not fix it.

You can read the resolution limit off a spec page. The next limit, the content filter, only shows up when a generation is refused.

## What does Seedance's content filter block, and how do you word around it?

Seedance's filter is predictable once you treat it as two gates: one on what the shot shows, one on whose face it uses. Moderation checks for sexual content at both the input and the output stage, on every platform. A separate face filter refuses the generation the moment a real person's photo is used as a reference. ByteDance's policy adds three more refusals: celebrity or public-figure likenesses, copyrighted characters, and graphic violence or gore without context.

![What Seedance's filters block: 1. Sexual content, at input and output 2. A real person's photo used as a reference 3. Celebrity or public-figure likeness 4. Copyrighted characters 5. Graphic violence or gore without context](https://uscale.ai/blog-media/seedance-2-5-product-videos/steps-what-seedance-s-filters-block-1200x360.svg)
For a product video, the face gate is the one that usually causes trouble. Launch films rarely need anything on the other lines, but they often need a person on screen. The usual instinct is to use a real one: the founder's headshot, a well-known tech figure, or "someone like" a famous CEO. Each of those runs into the face filter or the likeness rule. The same gate decides which portrait you can give the Talking Avatar model covered earlier.

### Describe the role, not the person

The prompts that get through describe a presenter by what the viewer should read from them, not by who they are. Instead of "a presenter who looks like [named executive]", write "a woman in her forties in a navy blazer, speaking calmly to camera in a bright office". That carries the tone you wanted from the famous name. It also gives the model features it can render instead of an identity it has to refuse. Characters work the same way: ask for "a friendly robot mascot with round blue eyes", never a named one from a film.

### The skip_moderation flag is not your workaround

The public Seedance API schema documents a skip_moderation parameter for both 2.0 and 2.5, and it defaults to false on both. So a standard request runs with moderation on. Write every prompt for the filter you will actually get.

Once your presenter brief clears the filter, the remaining question is which model should shoot it.

## Seedance vs. Kling vs. Veo: which fits a product video?

Seedance 2.5 costs the most per second of the three at 1080p. It earns its place through shot length and reference control, not price. Choose the shot the storyboard needs first, then check what that shot costs.

![Per-second cost at comparable top resolution: Seedance 2.5 (1080p, BytePlus) $0.5686/s, Kling 3.0 Pro (video input) $0.168/s, Veo 3.1 Standard (720p/1080p) $0.40/s](https://uscale.ai/blog-media/seedance-2-5-product-videos/chart-per-second-cost-at-comparable-top-resolu-1200x444.svg "Sources: ArgoLink / Seedance 2.5 API Pricing: 6 Providers Compared, Oct 6, 2026; Fastlane / Kling 3.0: Specs, Pricing and What It Means for Short-Form Marketers, Oct 1, 2026; WaveSpeed / Google Veo 3 Pricing 2026: API Costs and Access, Jan 1, 2026")
Kling 3.0's official API runs from about $0.084 per second in standard mode to $0.168 in Pro mode with video input. Veo 3.1 Standard on Google's Gemini API costs $0.40 per second at 720p or 1080p and $0.60 at 4K. Its cheaper Fast and Lite tiers sit below that. At 1080p, Seedance costs more than both.

### When each one wins

- **Seedance 2.5:** a long single-take dialogue scene, or a presenter clip from one portrait. It takes up to 30 images, 10 video clips and 10 audio clips per generation, and an audio track can drive lip-sync. Together these keep two speakers consistent without stitching.
- **Kling 3.0:** a short b-roll shot on a tight budget. Skip its free tier for a launch, because it is 720p, watermarked and barred from commercial use. The paid API is the real starting point. More in our [Kling 3.0 guide for product videos](https://uscale.ai/blog/kling-3-0-product-videos).
- **Veo 3.1:** a brief that calls for 4K. Seedance's public model generates natively at 1080p, and its separate 4K model is open only to selected enterprise trials.

The rule to reuse: price per second ranks the models, but the shot decides which ones can do the job at all. A cheaper model that can't hold the scene in one take means stitching and retakes.

Picking the right model is half the work. The next section covers the prompting mistakes that waste a generation.

## What mistakes do people make when prompting Seedance 2.5?

It's fair to expect less prep from Seedance 2.5. It renders a full 30-second clip in one generation, with no stitching or extension passes, and it accepts far more references than the version before it. Most failed first attempts don't come from the model's limits. They come from three wrong assumptions. Each one wastes a generation, and each has a fix you can apply before the next.

### Using a real colleague's photo as the presenter

The face filter rejects a generation as soon as a real person's photo is used as a reference. ByteDance's policy also bars real human faces in reference images outright. If you upload your CTO's headshot, you get a refusal instead of a talking CTO.

**Fix:** describe the presenter in words (age range, clothing, setting, manner) and let the model create a generic person.

### Treating the 30-second take as a substitute for a plan

One continuous shot removes the stitching. It does not remove the script. A 30-second take still has to fit every line, every pause and every turn in the scene, and the model cannot know where you wanted them.

**Fix:** write one beat per line with a rough time for each. Then give the model the voice track as an audio reference so the pacing and lip-sync follow it.

### Reading "1440p" as a sharper render

Seedance 2.5 renders natively at 1080p. The 1440p options some providers list, such as Atlas Cloud's 1440p-sr and 1440p-esr, take that 1080p source and scale it up to a 1440-pixel short edge. The frame gets bigger, but the render underneath is the same.

**Fix:** judge the clip at native 1080p. Pay for the upscale only if your delivery format needs the larger frame.

None of these is a flaw in the model. Each one comes from misreading the spec, and checking for them first is what makes a 30-second single shot worth its per-second rate.

So the shot pays off when the brief is right. What's left is turning that brief into a launch video people actually see.

## How to make your own product launch video with Seedance 2.5

Which route you take depends on what your launch needs, not on which model is newest. Three decisions get you there.

**Decide whether you need one shot or a whole film.** A single presenter line or a two-line dialogue scene is a shot. A launch video that shows your product working is a film. You're done when you can name which one you need in a sentence on your launch checklist.

**Write every line in a doc first.** This costs nothing and works for either route. Write one line per beat and give every line its own beat. It is also your quick win, something you can finish within the hour. If you do only one thing this week, do this, because both routes start from a script in a doc.

### If you need one shot

**Generate the shot yourself in Dreamina or Jimeng, or through the BytePlus API.** Describe a generic presenter instead of uploading a colleague's photo, because the face filter refuses real people as references. Record the product screens around the shot with the screen recorder you already have. You're done when the clip sits in your launch folder next to the screen recording.

### If you need a finished launch film

**Paste your product's site into uscale.** The director reads your product, asks what the film is for and proposes a form. That might be a [launch film](https://uscale.ai/use-cases/product-launch-video) or a founder avatar film made from one clear photo of you. Paste in the lines from your doc and ask the director to keep them word for word.

![The director in the uscale chat proposing a storyboard, beat by beat, with the narration line for each](https://uscale.ai/blog-media/how-to/chat-storyboard-800x1694.webp "Every narration line is approved here before anything is filmed — including a presenter clip.")
**Approve the storyboard.** Every narration line appears on it before anything is filmed. You're done when you've signed off each line, or asked in the chat to change the ones you don't want.

Success looks like the video on your launch page and one link that sales can send.

**Try it:** [make a product launch video](https://uscale.ai/use-cases/product-launch-video): a launch film of your real product, every narration line approved before filming.

## Before your next launch

The spec sheet matters less than the shot you actually need. Pick the step that matches where you are:

1. **If you need one presenter line or a short two-person exchange**, open Seedance 2.5 directly, in ByteDance's apps or through the API, and test a single shot before you budget around it.
2. **Write the lines first.** A long take still needs a beat for every line, and a generic presenter description keeps the filter out of your way.
3. **Price the clip at the resolution you'll ship**, not the one in the marketing copy.
4. **If you need a finished launch video**, with the right shot picked for each moment and the narration approved before anything is filmed, [start a product launch video in uscale](https://uscale.ai/use-cases/product-launch-video). Your first video is free, and no card is needed. The [example films](https://uscale.ai/examples) show what a finished one looks like, and [the credit packs](https://uscale.ai/pricing) cover the videos after that.

A good launch video isn't the newest model at its best. It's the right shot, on lines somebody agreed on, with the product doing what it does.

## Questions

### What is Seedance AI?

Seedance AI is ByteDance's AI video model, which turns text prompts, images, audio and video references into short video clips. Its current version, Seedance 2.5, had its stable release on July 31, 2026. It renders a full 30-second clip in a single generation, generates natively at 1080p, and accepts up to 50 reference inputs per generation, including audio tracks that can drive lip-sync in more than ten languages.

### Is Seedance AI free?

Seedance AI is not free through its official API. On BytePlus ModelArk, ByteDance's official platform, Seedance 2.5 is billed per second: $0.1028 at 480p, $0.2311 at 720p and $0.5686 at 1080p for requests without video input. Inside ByteDance's Jimeng and Dreamina apps, it is the default video model, and those apps set their own pricing. Every rendered second is billed, including takes you discard.

### How do you use Seedance AI?

You use Seedance AI in one of two places: ByteDance's Jimeng and Dreamina apps, where Seedance 2.5 is the default video model and no code is needed, or the Volcano Engine and BytePlus APIs, where you write each request and pay per second. For a presenter or dialogue shot, record the audio first, describe a generic person rather than uploading a real face, and test at 480p before rendering at 1080p.

### Who owns Seedance AI?

Seedance AI is developed by ByteDance. ByteDance released Seedance 2.5 on July 31, 2026, first as the default video model in its own Jimeng and Dreamina apps, then for developers through API access on Volcano Engine and BytePlus, which are ByteDance's developer platforms. In October 2026, ByteDance's Volcano Engine also released a separate Seedance 2.5 4K model for film and television production, open only to selected enterprise customers on trial.

### Is Seedance the best AI video generator?

Seedance is not the best for every job; it depends on the shot. Seedance 2.5 is strongest for long single-take dialogue scenes and presenter clips: it renders 30 seconds in one generation and lets an audio track drive lip-sync. It is also the priciest of three options at 1080p, at $0.5686 per second, against $0.40 for Veo 3.1 Standard and $0.084 to $0.168 for Kling 3.0. For 4K delivery, Veo 3.1 fits better; for cheap short b-roll, Kling.

---
Source: https://uscale.ai/blog/seedance-2-5-product-videos