Video generation is asynchronous: create a task (202 + task id), poll until the status is terminal, then read the signed URL. Same base URL and atp- key as image. The gateway routes your unified model to a video provider with same-name failover. Available video models: seedance-2-0 (standard), seedance-2-0-mini (lighter / lower cost), seedance-2-0-fast (faster rendering), and the Kling family (preview): kling-v3-standard/-pro, kling-o3-standard/-pro with -i2v / -reference / -reference-7 / -v2v / -video-edit variants — all billed per second by resolution (Kling audio generation not yet available); plus the Alibaba family (preview): wan-2-7-t2v, wan-2-7-i2v, happyhorse-1.1-t2v, happyhorse-1.1-i2v, happyhorse-1.1-r2v (up to 9 reference images) and happyhorse-1.0-video-edit (source clip 3–60 s + up to 5 reference images). HappyHorse duration is 3–15 s and its watermark defaults to on (pass watermark: false); wan-2-7-* currently renders and bills 1080P regardless of the requested resolution. Confirm names with GET /v1/models.
A succeeded task carries content.video_url — a signed edge URL with a 30-minute TTL. After expiry the task still reports succeeded but video_url is null and expired: true (re-create to regenerate).
Requests are refused with 402 insufficient_quota when the project balance is ≤ 0.
Check the model and project permission first
Call GET https://api.atptoken.ai/v1/models with the same key used to create the task. If a model is absent, that project cannot use it; installing a skill or knowing the model name does not bypass allowed models.
content[] blocks — Each block is { "type": "text" | "image_url" | "video_url" | "audio_url", ... }. Text only → text-to-video; include an image → image-to-video. image_url/video_url take { "url": "…" }; role is first_frame / last_frame / reference_image / reference_video.
What url accepts
Verified against production 2026-08-04
The video endpoint accepts public https:// URLs only.
Form
Video endpoint
Image endpoint
public https://… URL
works
works
data:image/…;base64,…
rejected — invalid_parameters: The parameter combination is not supported.
works
asset://<id> / asset://<pid>.<id>
not supported
not supported
asset:// was previously documented here as a valid form. It is not: both spellings, on both endpoints, fail at generation time with provider_error / generation_failed. Do not spend calls on it.
To reference something you uploaded, turn the upload into a public URL:
# 1. upload — the response key is `id` (an_<ULID>), not gw_file_id
curl -s https://api.atptoken.ai/v1/files \
-H "Authorization: Bearer atp-..." -F "file=@./first-frame.png"
# → 201 { "id": "an_01H...", "object": "file", "bytes": 152340, ... }
# 2. resolve to a no-auth URL — read the 302 Location, do NOT follow it
curl -sD - -o /dev/null https://api.atptoken.ai/v1/files/an_01H... \
-H "Authorization: Bearer atp-..." | grep -i '^location:'
# → location: https://<object-store>/gateway-files/...?<presigned> (~15-minute TTL)
Use that Location value as the url. It needs no authentication, which is what the upstream provider requires — but it expires in about 15 minutes, so resolve it immediately before creating the task rather than caching it.
wan-2-7-t2v / wan-2-7-i2v and the HappyHorse text/image/reference models use the same unified endpoint and content[] shape as every other video model — the notes below are the model-specific parts. happyhorse-1.0-video-edit is the exception: it is served on a separate DashScope-compatible endpoint (see below).
Model
Input
Duration
Resolution billed
Endpoint
wan-2-7-t2v
text
2–15 s
always 1080P (see caveat)
unified
wan-2-7-i2v
text + 1 first-frame image
2–15 s
always 1080P (see caveat)
unified
happyhorse-1.1-t2v
text
3–15 s
720P / 1080P as requested
unified
happyhorse-1.1-i2v
text + 1 first-frame image
3–15 s
720P / 1080P as requested
unified
happyhorse-1.1-r2v
text + 1–9 reference images
3–15 s
720P / 1080P as requested
unified
happyhorse-1.0-video-edit
1 source video (3–60 s) + up to 5 reference images
follows the source clip
720P / 1080P as requested
DashScope
wan-2-7 renders 1080P regardless of the requested resolution
wan-2-7-t2v / wan-2-7-i2v currently return a 1080P clip even when the request says 720P, and billing follows what was produced — so a 5-second clip costs the 1080P rate. Budget for 1080P, or use happyhorse-1.1-* when you need 720P pricing.
HappyHorse accepts 720P and 1080P only
These models support 720P and 1080P. A request for 480P is not rejected — it renders at roughly 1080P and is billed at the 1080P rate. Send 720P when you want 720P pricing.
Text-to-video
curl https://api.atptoken.ai/omni/media/v1/contents/generations/tasks \
-H "Authorization: Bearer atp-..." -H "Content-Type: application/json" \
-d '{
"model": "happyhorse-1.1-t2v",
"content": [{ "type": "text", "text": "a horse galloping across a green field" }],
"resolution": "720P", "ratio": "16:9", "duration": 5,
"watermark": false
}'
Image-to-video — add one image block with role: "first_frame":
Reference-to-video (happyhorse-1.1-r2v) — 1 to 9 reference images, each with role: "reference_image". Refer to them in the prompt as [Image 1], [Image 2], …:
Validate multi-image composition on a small run first
In one test against production (2026-08-04) a kling-o3-pro-reference task with two reference_image blocks and the prompt "place [Image 1] in the scene of [Image 2]" was accepted, rendered and billed — but the composition instruction was not carried out: the output kept Image 1's own background. That was a single test with synthetic source images, so it is not evidence the feature is broken. It does mean you should not assume [Image N] cross-image composition works: run one short, low-resolution clip and look at it before committing a batch.
Video editing (happyhorse-1.0-video-edit) — use the DashScope-compatible endpoint, not the unified one. The unified endpoint rejects this model with 400 video-edit requires a source video. Create the task with input.media[] and poll GET /omni/media/v1/tasks/{id}:
input.media[].type accepts video or reference_image only, and the URL key is url (not video_url). Every URL — source clip and reference images alike — must be publicly downloadable without authentication; if any of them cannot be fetched the task ends FAILED with Failed to download … and is not billed. This endpoint returns the DashScope response shape (output.task_status, output.video_url), and the output length currently follows the source clip rather than parameters.duration.
Billing — per second of output at the requested resolution, metered as video tokens (width × height × seconds × 24 fps ÷ 1024). For -r2v and -video-edit the input clip's seconds are billed too, so a 5-second edit of a 5-second source bills 10 seconds. Failed tasks are not billed. Rates are on the pricing page.
Other model-specific notes
Watermark: HappyHorse has the watermark on by default — pass watermark: false to disable it. Wan 2.7 defaults to off.
Ratio: HappyHorse also accepts 4:5, 5:4, 9:21, 21:9 on top of the shared list.
Prompt length: 5,000 characters (HappyHorse: 2,500 for Chinese text).
generate_audio is not available on these models yet.
