Wan 2.6 Reference to Video | Fast Image-to-Video API

WAN 2.6 Reference-to-Video

WAN 2.6 Reference-to-Video is ’s WanXiang 2.6 model for turning example videos + a text prompt into new shots. Provide up to two reference clips and the model learns their style, motion, and framing, then generates a new 5–10s video at up to 1080p.

🚀 Highlights

Reference-driven motion & style – Mimic camera moves, pacing and composition from your reference videos while following your prompt.
Up to two reference videos – Blend style from one clip and motion from another, or use different angles of the same scene.
Cinematic resolutions – Choose from 720p, or 1080p (portrait or landscape).
Story-aware generation – Works with prompt expansion and multishots to build richer, multi-shot sequences.
Audio-ready pipeline – Optional audio field for workflows that need motion aligned to external sound.

Output format: MP4 video at the selected size and duration.

🧩 Parameters

prompt* Text description of the new scene: characters, actions, environment, camera motion, mood, style, etc.
videos* 1–2 reference clips (URLs or uploads). These guide style, camera work, pacing, and motion structure.
negative_prompt Things to avoid, e.g. watermark, text, distortion, extra limbs.
audio (optional) External audio track for advanced pipelines where timing should loosely follow a given soundtrack. For most use cases you can leave this empty.
size One of the following resolution presets:
1280×720 or 720×1280 → 720p
1920×1080 or 1080×1920 → 1080p
duration Video length: 5 s or 10 s.
shot_type
single – Single-shot clip.
multi – When combined with enable_prompt_expansion, WAN 2.6 can break your idea into multiple shots of the same scene.
enable_prompt_expansion If enabled, ’s prompt optimizer expands short prompts into a richer internal script before generation.
seed Random seed. Set -1 for a new random result each time, or fix to a specific integer for reproducible layout and motion.

💰 Pricing

Resolution	Sizes (W×H)	5 s	10 s
720p	1280×720 / 720×1280	$1.00	$1.50
1080p	1920×1080 / 1080×1920	$1.50	$2.25

✅ How to Use

Prepare 1–2 reference videos

Clean motion, stable framing, and clear style work best.
You can use two angles of the same scene, or two stylistically similar clips.

Write your prompt

Describe what should happen in the new video, not just what’s in the references.
Example: “Cyberpunk alley at night, hero walking toward camera, slow dolly-in, neon reflections on wet ground, cinematic color grading.”

(Optional) Add a negative_prompt

Keep it short and focused: watermark, text, logo, extra limbs, low resolution.

Choose size and duration

720p/1080p according to your platform (Reels, TikTok, YouTube, etc.).
5 s for quick shots, 10 s for more complex actions.

Configure multishots & prompt expansion

Turn on enable_prompt_expansion for shorter prompts.
Enable multishots if you want WAN 2.6 to create a multi-shot sequence.

Set seed (optional)

Use a fixed seed to iterate while keeping composition similar.

Run the model and download the generated clip.

💡 Prompt & Reference Tips

Keep reference content and prompt aligned – if references show a city night scene, avoid asking for a sunny beach.
Use two references when you want to mix:
video A’s camera & motion + video B’s lighting/style.
Mention where you want the model to follow reference closely, e.g.: “Follow reference camera speed and angles, but change character outfit to futuristic armor.”
For portrait/vertical social content, select 480×832, 720×1280, or 1080×1920; for YouTube-style landscape, use the corresponding wide resolutions.

More Models to Try

vidu/reference-to-video-q2 Vidu’s Q2 reference-to-video model for turning style and motion from example clips into new shots, ideal for anime-style edits, trailers, and storyboards.
google/veo3.1/reference-to-video Google Veo 3.1 reference-conditioned video generator, designed for high-fidelity cinematic motion that closely follows your reference footage.
kwaivgi/kling-video-o1/reference-to-video Kwaivgi’s Kling Video O1 reference-to-video model, great for copying camera language and pacing from a sample clip while changing characters or scenes.
/seedance-v1-lite/reference-to-video SeeDance v1 Lite, a lightweight reference-to-video model for fast, style-consistent generations based on short example videos.

Wan 2.6 Reference To Video API — Quick start

Grab a WaveSpeedAI API key, then call POST https://api.wavespeed.ai/api/v3/alibaba/wan-2.6/reference-to-video with your input as JSON. The endpoint returns a prediction id; poll the prediction endpoint until status flips to completed, then read the output URL from data.outputs[0]. Examples for Wan 2.6 Reference To Video below.

HTTP example

# Submit the prediction
curl -X POST "https://api.wavespeed.ai/api/v3/alibaba/wan-2.6/reference-to-video" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $WAVESPEED_API_KEY" \
  -d '{
    "prompt": "A cinematic shot of a city at sunset, soft golden light",
    "negative_prompt": "blurry, low quality, distorted",
    "audio": "https://example.com/your-audio.mp3",
    "size": "1280*720",
    "duration": 5,
    "shot_type": "single",
    "enable_prompt_expansion": false,
    "seed": -1
}'

# Response includes a prediction id. Poll for the result:
curl -X GET "https://api.wavespeed.ai/api/v3/predictions/{request_id}/result" \
  -H "Authorization: Bearer $WAVESPEED_API_KEY"

# When status is "completed", read the output from data.outputs[0].

Node.js example

// npm install wavespeed
const WaveSpeed = require('wavespeed');

const client = new WaveSpeed(); // reads WAVESPEED_API_KEY from env

const result = await client.run("alibaba/wan-2.6/reference-to-video", {
        "prompt": "A cinematic shot of a city at sunset, soft golden light",
        "negative_prompt": "blurry, low quality, distorted",
        "audio": "https://example.com/your-audio.mp3",
        "size": "1280*720",
        "duration": 5,
        "shot_type": "single",
        "enable_prompt_expansion": false,
        "seed": -1
});

console.log(result.outputs[0]); // → URL of the generated output

Python example

# pip install wavespeed
import wavespeed

output = wavespeed.run(
    "alibaba/wan-2.6/reference-to-video",
    {
    "prompt": "A cinematic shot of a city at sunset, soft golden light",
    "negative_prompt": "blurry, low quality, distorted",
    "audio": "https://example.com/your-audio.mp3",
    "size": "1280*720",
    "duration": 5,
    "shot_type": "single",
    "enable_prompt_expansion": false,
    "seed": -1
}
)

print(output["outputs"][0])  # → URL of the generated output

Wan 2.6 Reference To Video API — Frequently asked questions

What is the Wan 2.6 Reference To Video API?

Wan 2.6 Reference To Video is a Alibaba model for video generation from images, exposed as a REST API on WaveSpeedAI. WAN 2.6 Reference-to-Video turns character, prop, or scene references—single or multi-view—into new video shots with preserved identity, style, and layout plus smooth, coherent motion. Ready-to-use REST inference API, best performance, no cold starts, affordable pricing. You can call it programmatically or try it from the playground above.

How do I call the Wan 2.6 Reference To Video API?

POST your input parameters to the model's REST endpoint (shown in the API tab of this playground) with your WaveSpeedAI API key in the Authorization header. Submission returns a prediction ID; poll the prediction endpoint until status flips to "completed", then read the output URL from the result. The playground generates a ready-to-paste code sample in Python, JavaScript, or cURL for whatever inputs you've set. Full request/response shape is documented at https://wavespeed.ai/docs/docs-api/alibaba/alibaba-wan-2.6-reference-to-video.

How much does Wan 2.6 Reference To Video cost per run?

Wan 2.6 Reference To Video starts at $0.50 per run. That figure is the base price — the final charge scales with the parameters you set in the form (output size, length, count, references, or whatever knobs this model exposes), so a higher-quality or larger output costs more than a minimal one. The exact cost for your current input is shown live next to the Generate button before you submit, and the actual per-call charge is recorded on the prediction afterwards.

What inputs does Wan 2.6 Reference To Video accept?

Key inputs: `prompt`, `audio`, `duration`, `size`, `seed`, `negative_prompt`. The full JSON schema (types, defaults, allowed values) is rendered above the Generate button and mirrored in the API reference at https://wavespeed.ai/docs/docs-api/alibaba/alibaba-wan-2.6-reference-to-video.

How long does Wan 2.6 Reference To Video take to generate?

Average end-to-end generation time on WaveSpeedAI is around 226 seconds per request — measured across recent runs. Queue time scales with global demand; live status is visible in the prediction record.

Can I use Wan 2.6 Reference To Video outputs commercially?

Commercial usage rights depend on the model's license, set by its provider (Alibaba). The license summary appears on the model card above; see WaveSpeedAI's Terms of Service for platform-level conditions.

EjemplosVer todo

Modelos relacionados

README