← All models
SCAIL-2 is an end-to-end character animation model that drives a reference character from a source video without relying on intermediate pose representations like skeleton maps.
Example
Details
- Accepts
- image + video
Pricing
- Price
- 12.5 cr / second
Prices in credits (1 credit = $0.01).
Data schema
Input
| Field | Type | Description |
|---|---|---|
| mode | stringenum: animation, replacement | `animation` animates the reference character with the driving motion. `replacement` replaces the driving subject with the reference character while keeping the original scene. |
| seed | — | Random seed. Leave empty for a random seed. The seed used is returned in the response. |
| shift | number | Flow-matching noise schedule shift. 3.0 is recommended for the 512p tier. |
| promptrequired | string | The prompt describing the final video to generate. |
| image_urlrequired | string | URL of the reference character image. The character in this image is animated (or used to replace a subject) according to the driving video. |
| video_urlrequired | string | URL of the driving (motion) video. The subject can be a human, multiple humans, or an animal. |
| resolution | stringenum: 512p, 704p | Output resolution. 512p outputs 896x512 (landscape) or 512x896 (portrait); 704p outputs 1280x704 or 704x1280. Orientation is chosen automatically from the reference image aspect ratio. |
| driving_type | stringenum: end_to_end, pose | Driving signal for animation mode. `end_to_end` (default) drives directly from the SAM3-masked driving video and is the recommended, most robust path. `pose` renders an explicit NLF/DWPose skeleton from the driving video (more control for challenging inputs). |
| subject_type | stringenum: human, animal | Type of subject in the driving video. `animal` switches the SAM3 detection prompt to non-human subjects. |
| guidance_scale | number | Classifier-free guidance scale. Controls prompt adherence versus creativity. |
| num_inference_steps | integer | Number of diffusion sampling steps. Higher improves quality but is slower. |
Output
| Field | Type | Description |
|---|---|---|
| seed | integer | The seed used to generate the video. |
| video | — | The generated video file. |
| preprocessing_video | — | Debug (only when `output_preprocessing=true`): the Stage-A pose-conditioning video fed to the diffusion model — the rendered NLF/DWPose skeleton for `driving_type=pose`, or the SAM3-masked driving video for `end_to_end`. |
| preprocessing_mask_video | — | Debug (only when `output_preprocessing=true`): the Stage-A driving mask video. |