Wan 2.7 Spicy Reference-to-Video API by Atlas Cloud
An image-reference-to-video model that generates one continuous video from one to four reference images with precise subject binding and prompt control. Supports image references only; reference video and audio are not accepted.
Wan 2.7 Spicy Reference-to-Video
atlascloud/wan-2.7-spicy/reference-to-video is an image-reference-to-video model. It turns one to four reference images and a text prompt into a continuous video, combining people, objects, and environments from separate images while following the requested action, composition, camera direction, and visual style.
Reference input scope
This model accepts images only as reference inputs. Unlike Wan 2.7 Reference-to-Video and Seedance 2.0 Reference-to-Video, which also support video and audio references, this model does not accept reference videos or reference audio.
Request
{ "model": "atlascloud/wan-2.7-spicy/reference-to-video", "reference_images": [ "https://static.atlascloud.ai/media/images/66c94270e84804c2029ef6095567c948.jpg", "https://static.atlascloud.ai/media/images/daef59ba6c4afe33225a455aee9e62b5.webp", "https://static.atlascloud.ai/media/images/23d7e0acfa10101f5819e43b42c0e2ff.jpg" ], "prompt": "A young woman@image1 is sitting on the sofa@image2, drinking a cocktail@image3.", "duration": 5, "resolution": "1440P-SR", "aspect_ratio": "16:9" }
Parameters
| Parameter | Required | Description |
|---|---|---|
model | Yes | Must be atlascloud/wan-2.7-spicy/reference-to-video. |
reference_images | Yes | Ordered array of 1тАУ4 image URLs or Base64 data URIs. Images are the only supported reference media type; video and audio references are not accepted. |
prompt | Yes | Describe the desired video and bind every reference with attached subject@imageN syntax. |
duration | No | Video duration from 2тАУ15 seconds. Default: 5. |
resolution | No | 720P, 1080P, 1080P-SR, or 1440P-SR. Default: 720P. |
aspect_ratio | No | auto, 16:9, 9:16, 4:3, 3:4, or 1:1. Default: auto. |
Reference images
Image numbering follows the order of reference_images:
reference_images[0] тЖТ image1 reference_images[1] тЖТ image2 reference_images[2] тЖТ image3 reference_images[3] тЖТ image4
Attach each image tag directly to the subject it represents:
young woman@image1 sofa@image2 cocktail@image3
Multiple subjects from the same image can share one tag:
woman in yellow@image1 and man in blue@image1
Every uploaded image must be referenced in the prompt, and every @imageN must correspond to an uploaded image.
Prompt guidance
Describe the visible action, environment, camera direction, composition, wardrobe, lighting, and style you want in the final video. Explicit instructions in the prompt are followed whenever they are visually compatible.
For best results:
- Identify each referenced subject clearly.
- Describe one continuous scene rather than multiple cuts or locations.
- State which subject should receive the most visual attention.
- Include any required camera angle, framing, or wardrobe details directly in the prompt.
Pricing
Pricing depends only on the selected resolution and billed video duration:
| Resolution | Price per second | Five-second price |
|---|---|---|
720P | $0.1200 | $0.6000 |
1080P | $0.1800 | $0.9000 |
1080P-SR | $0.1440 | $0.7200 |
1440P-SR | approximately $0.2560 | approximately $1.2800 |
Video generation is billed for at least five seconds. Reference-image count and aspect ratio do not change the price.
Asynchronous generation
Submit a request to /api/v1/model/generateVideo. The response contains a prediction ID that can be used to poll /api/v1/model/prediction/{request_id} until generation finishes. The final video URL is returned in data.outputs.
