Wan 2.2 Spicy API reference

Autenticación

Authorization: Bearer YOUR_API_KEY

Model & pricing

Billed on the duration you request (5 or 8 seconds), not on the exact length of the returned file. Adding a last frame costs nothing extra. Failed requests are not charged.

Limits (read this first)

Code examples

# ─── Image-to-video (first frame + prompt) ────────────────
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-2.2-i2v-spicy",
    "image": "https://example.com/first-frame.jpg",
    "prompt": "she slowly turns toward the camera, hair moving in the wind",
    "duration": 5,
    "resolution": "480p"
  }'
# Response (async — a task is created):
# { "code": 200, "data": { "taskId": "clxxx", "state": "pending" } }


# ─── First + last frame (the model interpolates between them) ──
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-2.2-i2v-spicy",
    "image": "https://example.com/first.jpg",
    "last_image": "https://example.com/last.jpg",
    "prompt": "smooth camera push-in",
    "duration": 8,
    "resolution": "720p",
    "seed": 42
  }'


# ─── Poll for the result ──────────────────────────────────
curl "https://api.apimodels.app/v1/video/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer YOUR_API_KEY"
# { "code": 200, "data": { "state": "completed", "resultUrls": ["https://r2.apimodels.app/..."] } }

Parámetros de la solicitud

Automatic prompt rewriting (prompt_extend)

When enabled, your prompt is first sent to a dedicated rewriting service that expands it and fills in detail, and the expanded version is what gets generated. It helps most with one-line prompts: if you write only "she turns her head", the rewrite adds camera, lighting and motion detail that the model responds to. The switch is on by default.

Pass prompt_extend: false to turn it off and have the model follow your text exactly. Turn it off when you have already written a long, precise prompt, or when you need reproducible comparisons alongside seed — otherwise each rewrite differs and a fixed seed still will not line up.

Note: there is no separate endpoint for the rewriting — this parameter is the only way to trigger it. It does not change what you pay: pricing depends only on resolution and duration.

What it is for: a cost-efficient traffic engine and interactive assets

Wan 2.2 Spicy is the volume model of the Spicy line: silent, 5 or 8 seconds, from $0.03/s at 480p. Silence is a feature here — with audio and video decoupled, your engine controls TTS, lip sync and background music precisely. For premium clips with native audio, 2 to 15 seconds and 1080p, use wan-2.7-i2v-spicy.

The latest upstream build cut 2.2 sampling time roughly in half: about 50 seconds for a 480p 5-second clip. The 35 to 75 second range in the limits table above is our own measurement from 09-20; expect the lower end.

Notas

• Async: create a task for a taskId, then poll GET ?task_id= for the result
• The first frame takes a public image URL or base64; the last frame is optional and makes the model interpolate between the two
• The model page has a web playground with the same parameters
• Result files are deleted after 30 days; download or re-host them in time
• Charged only on success