# Seedance

Source: https://docs.prodia.com/models/seedance/

Seedance is ByteDance's video generation model family, developed by the team behind Doubao and the [Seedream](https://docs.prodia.com/models/seedream/) image models. It is built for fast, prompt-faithful video generation: a single short prompt produces a coherent 5- or 10-second clip at up to 1080p, with strong adherence to camera-direction language ("dolly in", "tracking shot", "static") and good motion stability across frames.

### Architecture

Seedance is a Diffusion Transformer (DiT) trained jointly on text-to-video and image-to-video objectives. Two design choices stand out for API users:

- *Decoupled spatial and temporal layers* — the model alternates spatial-attention blocks (per-frame composition) with temporal-attention blocks (cross-frame motion). This split lets it learn fine-grained appearance and physics independently, which produces fewer flicker artifacts on textures like fur, fabric, and water than single-stream architectures.
- *Multi-shot training* — Seedance is trained on multi-shot sequences, so a single prompt that describes a brief sequence of actions ("the dog runs to the camera, then leaps over the puddle") tends to produce a continuous shot that follows the described beats rather than a single static motion.

### Variant comparison

The Seedance family on Prodia ships in two variants. *Pro Turbo* is the default starting point for almost all use cases; *Pro* exists for cases where you need maximum quality and are willing to pay more for it.

| Feature                 | Pro Turbo                       | Pro                             |
| ----------------------- | ------------------------------- | ------------------------------- |
| Job type prefix         | `inference.seedance.proturbo.*` | `inference.seedance.pro.*`      |
| Resolutions             | 480p, 1080p                     | 480p, 1080p                     |
| Aspect ratios (txt2vid) | 16:9, 9:16, 1:1, 4:3, 3:4, 21:9 | 16:9, 9:16, 1:1, 4:3, 3:4, 21:9 |
| Duration                | 5s or 10s                       | 5s or 10s                       |
| Generation time         | \~45s                           | \~60s                           |
| Cost (1080p, 5s)        | lower (around half of Pro)      | higher                          |

*Pro Turbo* — the recommended default. A distillation/optimization tier with the same API surface as Pro and \~25% faster generation. The quality difference for typical prompts is small enough that Pro Turbo is the right choice unless you have a specific reason to pay more.

*Pro* — full-quality variant. Useful when you have a finished prompt that didn't quite render the way you wanted on Pro Turbo and you want one more pass with the higher-quality model before iterating on the prompt.

> note:
>
> A `lite` variant (`inference.seedance.lite.*`) is also exposed, but it is *deprecated* and reaches end of life on 2026-04-30. New projects should use `proturbo` or `pro`. The migration is a one-token rename in the `type` field — the config schema is compatible.

### When to use Seedance

| Use case                              | Job type                                     | Why Seedance                                                 |
| ------------------------------------- | -------------------------------------------- | ------------------------------------------------------------ |
| Short cinematic clip from a prompt    | `proturbo.txt2vid`                           | \~45s generation at 1080p, strong camera-direction prompting |
| Animating a hero image / product shot | `proturbo.img2vid`                           | Faithful starting frame + natural motion in 5s               |
| Loopable social content               | `proturbo.txt2vid` with `aspect_ratio: 9:16` | Native portrait support for TikTok/Reels                     |
| Final-pass quality on a locked prompt | `pro.txt2vid` / `pro.img2vid`                | Slightly higher fidelity than Pro Turbo at higher cost       |

For *audio-driven* video (lip-sync from a portrait + audio file) or *first-and-last-frame* interpolation, [Wan 2.7](https://docs.prodia.com/models/wan-2-7/) is the better fit — Seedance does not accept audio or a `last_frame`. For very low-latency 720p-only generation, [Wan 2.2 Lightning](https://docs.prodia.com/models/wan-2-2/) generates in \~22s. For the most precise *camera choreography* (per-axis pan, tilt, roll, zoom), [Kling](https://docs.prodia.com/models/kling/) exposes those controls explicitly.

### Job types

| Job type                                 | Description                                     | ETA   |
| ---------------------------------------- | ----------------------------------------------- | ----- |
| `inference.seedance.proturbo.txt2vid.v1` | Text-to-video, Pro Turbo                        | \~45s |
| `inference.seedance.proturbo.img2vid.v1` | Image-to-video from a starting frame, Pro Turbo | \~45s |
| `inference.seedance.pro.txt2vid.v1`      | Text-to-video, Pro                              | \~60s |
| `inference.seedance.pro.img2vid.v1`      | Image-to-video from a starting frame, Pro       | \~60s |

> note:
>
> All Seedance job types are *async-only*. Submit to `POST /v2/job/async` and poll `GET /v2/job/async/:id/job.state.current` until it returns `processed`, then download `output/video.mp4`. See the [async reference](https://docs.prodia.com/reference/async/) and the [polling async jobs guide](https://docs.prodia.com/guides/polling-async-jobs/) for the full flow.

### Parameters

*Common to all job types:*

- `prompt` (required) — text description of the desired clip, 3–4,096 characters
- `resolution` — `"480p"` or `"1080p"` (default: `"1080p"`). Exact width/height varies with aspect ratio.
- `duration` — `5` or `10` seconds (default: `5`)
- `camera_fixed` — `true` to discourage camera motion (default: `false`, not guaranteed)
- `watermark` — `true` to add an "AI-generated" watermark (default: `false`)
- `seed` — integer for reproducible results

*Text-to-video only:*

- `aspect_ratio` — `"16:9"` (default), `"9:16"`, `"1:1"`, `"4:3"`, `"3:4"`, or `"21:9"`

*Image-to-video only:*

- `first_frame` — filename of the starting frame, supplied as a multipart form file. Aspect ratio of the output follows the input image.

### Prompting tips

- *Describe the motion, not just the scene* — Seedance is trained to follow camera and subject-motion language. A prompt like `"slow dolly in on the puppy, shallow depth of field"` produces noticeably more intentional motion than `"a puppy"`.
- *Use camera\_fixed for product shots* — when you want the subject to move but the camera held still (e.g. a rotating product on a tabletop), set `camera_fixed: true` and describe the subject motion in the prompt.
- *Match aspect ratio to platform* — `9:16` for TikTok/Reels/Shorts, `16:9` for YouTube/landing pages, `1:1` for Instagram feed, `21:9` for cinematic banners.
- *480p is genuinely useful for iteration* — 480p costs roughly a fifth of 1080p and finishes in the same wall time. Iterate on prompts at 480p, then switch the same prompt to 1080p for the final render.
- *Short, declarative prompts beat long ones* — Seedance handles 1–2 sentence prompts well. Very long prompts tend to get partially ignored; if you need to describe a sequence of actions, separate beats with commas rather than long subordinate clauses.

### Examples

Text-to-video at 1080p, 16:9, 5 seconds:

```json
{
  "type": "inference.seedance.proturbo.txt2vid.v1",
  "config": {
    "prompt": "a golden retriever puppy running through a field of wildflowers in slow motion, cinematic lighting, hyperrealistic",
    "resolution": "1080p",
    "aspect_ratio": "16:9",
    "duration": 5
  }
}
```

Sample frame from the generated clip (480p preview shown for size):

![Seedance Pro Turbo text-to-video output, golden retriever puppy](https://docs.prodia.com/llms/assets/c52c6edb2b55-seedance-txt2vid-poster.jpg)

Vertical 9:16 clip for TikTok/Reels with a fixed camera:

```json
{
  "type": "inference.seedance.proturbo.txt2vid.v1",
  "config": {
    "prompt": "a barista pouring latte art into a cup, steam rising, warm morning light",
    "resolution": "1080p",
    "aspect_ratio": "9:16",
    "duration": 5,
    "camera_fixed": true
  }
}
```

Image-to-video — animate a still frame with a described camera move. The `first_frame` field names the file, which is uploaded alongside the job as multipart form-data:

```json title="job.json"
{
  "type": "inference.seedance.proturbo.img2vid.v1",
  "config": {
    "first_frame": "cyberpunk-input.jpg",
    "prompt": "the camera slowly pushes forward through the cyberpunk market, neon signs flicker, people walk by",
    "resolution": "1080p",
    "duration": 5
  }
}
```

Input frame:

![Seedance img2vid input — cyberpunk street](https://docs.prodia.com/llms/assets/10e1e0e73d20-seedance-img2vid-input.jpg)

Sample frame from the generated clip — the camera has pushed forward through the same scene:

![Seedance Pro Turbo img2vid output frame](https://docs.prodia.com/llms/assets/1e4e2a624f53-seedance-img2vid-poster.jpg)

### Guides

[Generating Videos](https://docs.prodia.com/guides/generating-videos/) — Step-by-step tutorial covering text-to-video and image-to-video against the v2 inference API.

[Polling Async Jobs](https://docs.prodia.com/guides/polling-async-jobs/) — The submit → poll → download flow used by all Seedance job types.

[Async API Reference](https://docs.prodia.com/reference/async/) — Endpoint reference for /v2/job/async.