# Wan 2.7

Source: https://docs.prodia.com/models/wan-2-7/

Wan 2.7 is the most capable release in Alibaba's Wan model family — a single unified model that handles five generation tasks: text-to-image, image-to-image editing, text-to-video, image-to-video animation, and video-to-video continuation. It is the go-to choice when you need maximum flexibility and quality across both image and video modalities.

### Architecture

Wan 2.7 builds on the DiT (Diffusion Transformer) foundation established in the Wan 2.x series with a refined Flow-Matching framework. Key architectural improvements over earlier versions include:

- *Improved motion coherence* — reduced temporal flickering on skin, fabric, and moving objects through better physics-based motion consistency
- *Bilingual T5 encoder* — native support for both Chinese and English prompts with cross-attention in every transformer block
- *Prompt extension* — an intelligent prompt rewriting system that expands short prompts into detailed scene descriptions for better results (enabled by default, can be disabled)
- *Thinking mode* — for image generation, an optional reasoning step that improves composition and detail at the cost of longer generation time

### What sets Wan 2.7 apart

Wan 2.7 is uniquely versatile among the models on Prodia:

- *Five generation modes in one model* — no other model covers txt2img, img2img, txt2vid, img2vid, and vid2vid in a single family
- *Audio-driven video* — provide a WAV or MP3 file to drive lip-sync and motion timing in img2vid, useful for talking-head videos and music-driven content
- *First and last frame control* — specify both the starting and ending frames of a video, and the model generates everything in between. This enables loopable videos and precise scene transitions
- *Video continuation* — extend an existing video clip with vid2vid, maintaining visual consistency while adding new content
- *1080p video output* — one of the few open video models supporting full HD generation
- *Up to 15-second clips* — longer durations than most competing video models

### When to use Wan 2.7

| Use case                      | Job type                    | Why Wan 2.7                                                       |
| ----------------------------- | --------------------------- | ----------------------------------------------------------------- |
| High-quality image generation | `txt2img`                   | Thinking mode produces detailed, well-composed images at up to 2K |
| Image style transfer          | `img2img`                   | Edit or restyle photos with bilingual prompts                     |
| Cinematic video from text     | `txt2vid`                   | 1080p output with 5 aspect ratios and up to 15s duration          |
| Talking-head videos           | `img2vid` with `audio`      | Audio-driven lip-sync from a portrait photo                       |
| Product animations            | `img2vid`                   | Animate product shots with controlled motion                      |
| Loopable social content       | `img2vid` with `last_frame` | Set first = last frame for seamless loops                         |
| Scene extensions              | `vid2vid`                   | Continue an existing clip naturally                               |

For faster video generation at 720p where you don't need audio, frame control, or 1080p — [Wan 2.2 Lightning](https://docs.prodia.com/models/wan-2-2/) generates in \~22 seconds vs Wan 2.7's \~200 seconds.

### Job types

| Job type                      | Description                       | ETA    |
| ----------------------------- | --------------------------------- | ------ |
| `inference.wan2-7.txt2img.v1` | Generate an image from text       | \~40s  |
| `inference.wan2-7.img2img.v1` | Edit or restyle an existing image | \~40s  |
| `inference.wan2-7.txt2vid.v1` | Generate a video from text        | \~200s |
| `inference.wan2-7.img2vid.v1` | Animate an image into a video     | \~200s |
| `inference.wan2-7.vid2vid.v1` | Continue or extend a video clip   | \~200s |

> note:
>
> Video job types (`txt2vid`, `img2vid`, `vid2vid`) support async processing. See the [async reference](https://docs.prodia.com/reference/async/) for polling-based workflows.

### Parameters

*Common to all job types:*

- `prompt` — text description, up to 5,000 characters (required for most types)
- `seed` — integer 0–2147483647 for reproducible results

*Image generation (txt2img, img2img):*

- `size` — `1K` (\~1024x1024) or `2K` (\~2048x2048, default)
- `thinking_mode` — enable reasoning for improved composition (txt2img only, default: true)
- `image` — input image filename (img2img only)

*Video generation (txt2vid, img2vid, vid2vid):*

- `resolution` — `720P` (default) or `1080P`
- `ratio` — aspect ratio: `16:9` (default), `9:16`, `1:1`, `4:3`, `3:4` (txt2vid only)
- `duration` — video length in seconds, 2–15 (default: 5)
- `negative_prompt` — content to exclude, up to 500 characters
- `prompt_extend` — intelligent prompt rewriting (default: true)

*Image-to-video additional parameters:*

- `image` — first-frame image to animate
- `last_frame` — target last-frame image for start-end interpolation
- `audio` — driving audio for lip-sync and motion timing (WAV/MP3, 2–30s, max 15 MB)

*Video-to-video additional parameters:*

- `video` — input video clip to continue (MP4/MOV, 2–10s, max 100 MB)
- `last_frame` — target last-frame image to guide the continuation endpoint

### Prompting tips

- *Use prompt extension:* leave `prompt_extend` enabled (default) for short prompts — the model will expand them into detailed scene descriptions that produce better results
- *Negative prompts matter for video:* adding `"low resolution, error, worst quality, deformed"` as a `negative_prompt` noticeably improves video quality
- *Bilingual prompts:* you can mix Chinese and English in the same prompt for nuanced descriptions
- *Aspect ratios:* match your target platform — `9:16` for TikTok/Reels, `16:9` for YouTube, `1:1` for Instagram posts
- *Audio-driven video:* for best lip-sync results, use clear speech audio without background music. Audio longer than the video duration is automatically trimmed

### Examples

Text-to-video at 1080p:

```json
{
  "type": "inference.wan2-7.txt2vid.v1",
  "config": {
    "prompt": "A kitten running in the moonlight",
    "resolution": "1080P",
    "ratio": "16:9",
    "duration": 5,
    "negative_prompt": "low resolution, error, worst quality, deformed"
  }
}
```

Audio-driven talking head:

```json
{
  "type": "inference.wan2-7.img2vid.v1",
  "config": {
    "image": "portrait.jpg",
    "audio": "speech.mp3",
    "prompt": "A person speaking naturally, subtle head movement",
    "resolution": "720P",
    "duration": 10
  }
}
```

Text-to-image with thinking mode:

```json
{
  "type": "inference.wan2-7.txt2img.v1",
  "config": {
    "prompt": "A serene mountain landscape at sunrise with vibrant colors",
    "size": "2K",
    "thinking_mode": true
  }
}
```

### Guides

[Generating Images](https://docs.prodia.com/guides/generating-images/) — Step-by-step guide for generating images with Prodia.

[Generating Videos](https://docs.prodia.com/guides/generating-videos/) — Step-by-step guide for text-to-video and image-to-video generation.