Migrating Seedance from fal or ModelArk to Dawnwood Models

In shortFrom BytePlus ModelArk, change two lines — the base URL to https://models.dawnwood.ai/v1 and the key to a Dawnwood token — because Dawnwood Models accepts the ModelArk request body, path and task object as they are. From fal, the request is reshaped once - the prompt and image URLs become items in a content array, aspect_ratio becomes ratio, duration becomes an integer - and polling moves from fal's status and response URLs to one GET on the task id. Both paths are shown side by side below.

Before you start

1. Create an account at console.dawnwood.ai and open Developers.
2. Mint an API key. It is a Dawnwood personal access token (dw_pat_v1_…) with the model_api scope; it is shown once.
3. Buy credits. The API spends purchased credits only — signup credits and subscription-included credits are for the production pipeline.

From BytePlus ModelArk: two lines

Dawnwood Models serves the ModelArk content-generation task path as an alias, takes the same body (model, content[], resolution, ratio, duration, generate_audio) and returns the same task object with the same status values. The official dreamina-seedance-… model ids are accepted, and so are the short ids (seedance-2.0-fast, seedance-2.0-mini, seedance-2.0, seedance-2.5).

# before
BASE = "https://ark.ap-southeast.bytepluses.com/api/v3"
KEY = os.environ["ARK_API_KEY"]

# after
BASE = "https://models.dawnwood.ai/v1"
KEY = os.environ["DAWNWOOD_API_KEY"]   # dw_pat_v1_...

# unchanged
H = {"Authorization": f"Bearer {KEY}"}
task = requests.post(f"{BASE}/contents/generations/tasks", headers=H, json=body).json()
status = requests.get(f"{BASE}/contents/generations/tasks/{task['id']}", headers=H).json()

Two differences to know about. Video reference inputs (video_url items) are not supported yet and return 400 unsupported_content. And 1080p is not offered yet; requests use 480p or 720p.

From fal: reshape the request once

fal exposes each Seedance variant as its own endpoint and takes a flat body. Dawnwood Models has one endpoint, selects the model with a field, and takes inputs as a typed content array.

# before: fal queue API
r = requests.post("https://queue.fal.run/bytedance/seedance-2.0/reference-to-video",
                  headers={"Authorization": f"Key {os.environ['FAL_KEY']}"},
                  json={"prompt": prompt,
                        "image_urls": [lead_png, room_png],
                        "resolution": "720p", "aspect_ratio": "9:16", "duration": "5"}).json()
# poll r["status_url"] until COMPLETED, then GET r["response_url"] -> ["video"]["url"]

# after: Dawnwood Models
H = {"Authorization": f"Bearer {os.environ['DAWNWOOD_API_KEY']}"}
task = requests.post("https://models.dawnwood.ai/v1/video/generations", headers=H, json={
    "model": "seedance-2.0",
    "content": [
        {"type": "text", "text": prompt},
        {"type": "image_url", "image_url": {"url": lead_png}, "role": "reference_image"},
        {"type": "image_url", "image_url": {"url": room_png}, "role": "reference_image"},
    ],
    "resolution": "720p", "ratio": "9:16", "duration": 5,
}).json()
# poll GET /v1/video/generations/{id} until status is succeeded -> ["content"]["video_url"]

Field mapping

Prompt syntax that refers to reference images by position carries over, because the images keep their order in content[].

After the switch

• Retries. Send an Idempotency-Key header with each create call; a retried request returns the same task and is never charged twice.
• Concurrency. Each key has a limit on running tasks. Past it you get 429 with Retry-After; wait and resubmit.
• Failures. A failed or upstream-rejected task is refunded in full; the task's billing.refunded is true.
• Errors. Error bodies follow ModelArk's shape, {"error": {"code": …, "message": …}}, so existing error handling keeps working.

The full schema is in the API reference and as an OpenAPI 3.1 file at dawnwood.ai/models/openapi.json. Prices per model are on the pricing comparison.