Reference pictures
Make a clip that keeps the look of pictures you give, without fixing any one frame.
- Pick a model that takes them.
- Say what must stay the same.
- Send the job with the pictures in
input_references. - Wait and fetch the clip.
1. Pick a model that takes them
No field of the video models list says whether a model takes reference pictures. Its description does, and on one model its price. On 2026-09-29, these models named them.
| Lab | Models | The list says |
|---|---|---|
| Alibaba | alibaba/wan-3.0 | reference-guided video generation |
| Alibaba | alibaba/wan-2.7 | multiple reference images guide the style and content |
| Alibaba | alibaba/happyhorse-1.0, alibaba/happyhorse-1.1 | a set of reference images |
| ByteDance | bytedance/seedance-2.0, -2.0-fast, -2.0-mini | multimodal reference-to-video |
| ByteDance | bytedance/seedance-2.5 | multimodal reference-based generation |
| xAI | x-ai/grok-imagine-video | reference-conditioned |
| MiniMax | minimax/hailuo-2.3 | reference images as input |
| MiniMax | minimax/hailuo-3 | its pricing_skus holds reference_images |
Read a model's description on its page in Models before you build on it. The other fields that narrow the choice: Choose a video model.
2. Say what must stay the same
The pictures show the model what to keep. The prompt says what to keep, and what to change around it.
| Prompt part | Example |
|---|---|
| What the pictures hold | the mark from the reference picture |
| What must not change | its shape and its orange gradient |
| The new scene | carved in a clay tile on a potter's bench in Maatkas |
| Camera and light | slow orbit, soft window light |
3. Send the job
The sample sends the OpenDunes mark as the one reference picture, to bytedance/seedance-2.0-fast, for 5 seconds at 720p in 1:1. Up to the model, more pictures of the same thing, from other angles, hold it steadier.
import json
import os
import urllib.request
API_KEY = os.environ["OPENDUNES_API_KEY"]
ADDRESS = "https://opendunes.com/v1/videos"
MODEL = "bytedance/seedance-2.0-fast"
REFERENCE = "https://opendunes.com/icon.png"
PROMPT = (
"The mark from the reference picture, carved in a clay tile on a "
"potter's bench in Maatkas, slow orbit, soft window light. Keep the "
"mark's shape and its orange gradient."
)
request = urllib.request.Request(
ADDRESS,
data=json.dumps(
{
"model": MODEL,
"prompt": PROMPT,
"duration": 5,
"resolution": "720p",
"aspect_ratio": "1:1",
"input_references": [
{"type": "image_url", "image_url": {"url": REFERENCE}},
],
}
).encode(),
headers={
"Authorization": "Bearer " + API_KEY,
"Content-Type": "application/json",
},
)
with urllib.request.urlopen(request) as response:
job = json.load(response)
print(job["id"], job["status"])In each input_references entry | Holds |
|---|---|
type | image_url |
image_url.url | the picture's address: public HTTPS, answering with the picture itself |
A picture behind a login, a cookie or a web page is not a picture the provider can read. Test the address in a private browser window: it must show the picture alone.
The answer is 202, with the job's id and status pending. Every field: Clips.
References guide the look; they fix no frame. To make a picture the clip's exact first or last frame, send it in frame_images: Picture to video. When both are sent, frame_images wins.
4. Wait and fetch the clip
The job runs as any other: ask GET /v1/videos/{id} every ten seconds until it ends, then fetch the clip from /v1/videos/{id}/content?index=0, or give a callback_url and have the end posted to you. The whole loop, in four languages: Text to video. The post: Video webhooks.
The clip is charged in DADAThe Algerian dinar, the one currency the docs show.More on Balance and charges → once, when the job reports completed. On some models each picture sent in is charged too: the model's page in Models lists what it charges for, and Choose a video model names the models. A job that ends any other way charges nothing.
Last modified on