image to video API models: which ones accept a start frame
·
Three video adapters pass a first reference URL from image_inputs to the
upstream request. This is an adapter contract, not a guarantee that every
upstream endpoint honors image conditioning. Confirm the current upstream
model contract and inspect the output before relying on reference fidelity.
Image-generation adapters return still images and are not image-to-video models.
| Model | Cost | Duration | Resolution |
|---|---|---|---|
| Seedance | 25 credits | 3–15s | 720p, 1080p |
| Kling | 35 credits | 5 or 10s | 1080p |
| Veo 3 | 100 credits | 8s fixed | 720p, 1080p |
The request
{
"model": "seedance-2.0-fast",
"prompt": "the camera pushes slowly forward",
"image_inputs": ["https://example.com/frame.jpg"]
}A URL, not an upload — the image must be reachable by the upstream provider. Only the first entry is used; a second URL does not produce an interpolation between two frames.
Price does not change
The local pricing functions do not add a reference-image surcharge. The table shows the default request estimate; changing duration or resolution may change the held credits. For example, Veo 3 uses 50 credits at 720p instead of the default 100 credits at 1080p. This does not establish upstream input support.
The prompt should describe motion
The frame already contains the subject. Re-describing it is the most common reason the result comes back looking frozen — spend the words on camera movement and what changes. The image to video prompt generator has no subject field for that reason.
Continue with choose a model
Models covered on this page
Where these facts come from
- codebase: src/ai/providers/*.ts — every pricing() and paramsSchema
- codebase: src/config/website.tsx — credit pack pricing