Wan2.7 Reference-to-Video

POST /api/v1/services/aigc/video-generation/video-synthesis — Wan2.7 reference-to-video (identity / voice consistency)

The Wan 2.7 reference-to-video model accepts multimodal references (images, video, audio) and produces a video that keeps subject identity and voice consistent — single-character performance or multi-character interaction.

Models

Model	Notes
`wan2.7-r2v`	Wan 2.7 reference-to-video

Create a job

POST/api/v1/services/aigc/video-generation/video-synthesis

curl https://api.modelsite.ai/api/v1/services/aigc/video-generation/video-synthesis \
  -H "Authorization: Bearer $MODELSITE_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "model": "wan2.7-r2v",
    "input": {
      "prompt": "The character from Video 1 holds Figure 2 and plays a gentle folk tune; Figure 1 walks past and puts down what she carries",
      "media": [
        { "type": "reference_image", "url": "https://example.com/girl.jpg",
          "reference_voice": "https://example.com/girl-voice.mp3" },
        { "type": "reference_video", "url": "https://example.com/role2.mp4" },
        { "type": "reference_image", "url": "https://example.com/prop.png" }
      ]
    },
    "parameters": { "resolution": "720P", "ratio": "16:9", "duration": 10 }
  }'

Request body

Field	Type	Notes
`model`	string (required)	`wan2.7-r2v`
`input.prompt`	string (required)	≤5000 characters. Refer to media entries as "Figure 1 / Video 1" (images and videos count separately, in array order); with a single asset, "the reference image" also works
`input.negative_prompt`	string (optional)	≤500 characters (in `input`)
`input.media`	array (required)	`reference_image` + `reference_video` together 1–5 entries; optionally plus at most 1 `first_frame` for joint control
`parameters`	object (optional)	Below

media entry fields

Field	Notes
`type`	`reference_image` (subject/scene) / `reference_video` (subject + voice; avoid empty shots) / `first_frame` (joint control, ≤1)
`url`	Asset URL or Base64. Images ≤20MB, sides [240, 8000]px; video mp4/mov, 1–30 s, ≤100MB
`reference_voice`	Only on reference entries: a voice-reference audio URL (wav/mp3, 1–10 s, ≤15MB). Voice only, independent of content; takes precedence over the source audio of a reference_video

parameters

Parameter	Type	Notes
`resolution`	string	`720P` / `1080P` (default). Sellable tiers follow the price configuration
`ratio`	string	`16:9` (default) / `9:16` / `1:1` / `4:3` / `3:4`. Ignored when `first_frame` is present (follows the frame)
`duration`	integer	Default 5. `[2, 10]` when a reference video is present, `[2, 15]` otherwise
`prompt_extend`	boolean	Default `true`
`watermark`	boolean	Default `false`
`seed`	integer	`[0, 2147483647]`

Keep a single character per reference asset. Billed duration = input video duration + output duration (the poll's usage.duration).

Polling and result

GET/api/v1/tasks/{task_id}

curl "https://api.modelsite.ai/api/v1/tasks/$TASK_ID" \
  -H "Authorization: Bearer $MODELSITE_API_KEY"

Response fields

Field	Meaning
`output.task_id`	Job ID (same value returned at creation)
`output.task_status`	`PENDING` queued / `RUNNING` processing / `SUCCEEDED` done / `FAILED` failed
`output.video_url`	The generated video URL — only on `SUCCEEDED`; download promptly
`output.code` / `output.message`	Present only on failure, with the reason
`usage`	Usage stats (duration, resolution tier, …); counted only on success
`request_id`	Unique request ID — include it when reporting issues

Status flow: PENDING → RUNNING → SUCCEEDED / FAILED.

Billing

Billed by actual generated seconds — with a reference video, its duration counts too; failed jobs are not billed. Rates on the Models page.

Error handling

Creation-time errors return {"code": "...", "message": "..."}; mid-job failures surface through the poll's output.code / output.message. General error codes: Errors.