Introduction
Wan 2.6 Flash Reference to Video casts people, animals or objects from your references into a new scene you describe, on the faster, lower-cost Flash tier of Wan 2.6. Send up to five references, up to three of them short clips and the rest still images, or stills alone; location, action, camera and light all come from the prompt. The result runs any whole number of seconds from 2 to 10, at 720p or 1080p, in one of five aspect ratios, as one continuous take or a multi-shot edit. Clips are silent by default; switch generate_audio on when the scene needs its own soundtrack.
Choose a Wan 2.6 Flash mode by starting material
| What you have at the start | Mode |
|---|---|
| an approved opening frame | Wan 2.6 Flash Image to video |
| photos or short clips of the people and objects to cast | Wan 2.6 Flash Reference to video |
Key features & highlights
Cast from stills, clips, or both
reference_image_urls takes up to five stills and reference_video_urls up to three short clips, five references in total. Stills alone are enough: two portraits are all it takes to put two people in the same scene, and still images are not billed. A clip carries what a photograph cannot, such as how someone moves or how a place behaves, at the cost of its seconds being billed with the output.
The prompt names the cast
Wan identifies each reference only through the names character1, character2 and so on, in order: clips first, then stills. Name each one before describing what they do, for example character1 dancing with character2, and the model knows who is who. Everything else — the set, the action, the camera and the light — is written, so a person photographed once can be placed on a rooftop at night or a rain-soaked quay without a second shoot.
Sound as a switch, shape as a parameter
generate_audio is off by default and returns picture only; on, the clip comes back with generated voices, effects and ambience at twice the per-second rate. aspect_ratio is chosen per request: 16:9, 9:16, 1:1, 4:3 or 3:4 at either resolution. 720p returns 1280×720, 720×1280, 960×960, 1088×832 or 832×1088; 1080p returns 1920×1080, 1080×1920, 1440×1440, 1632×1248 or 1248×1632. Nothing is inherited from the references, so a vertical delivery does not need vertical sources.
Request parameters
| Parameter | Required | Description |
|---|---|---|
prompt | Yes | The new scene: name each reference as character1, character2 and so on, clips first and then stills, and say what each one does, then the setting, camera and light. at least 1 characters · at most 1500 characters |
negative_prompt | No | What must not appear: extra people, captions, a prop from a reference you do not want carried over. Up to 500 characters. at most 500 characters |
reference_video_urls | No | Up to three short clips of the subjects or places to cast, adding up to five seconds at most. They are numbered first, as character1, character2 and character3 in upload order. at least 1 items · at most 3 items |
reference_image_urls | No | Up to five still images of the people, animals or objects to cast, alongside the clips or instead of them; with the clips a request takes five references at most. at least 1 items · at most 5 items |
aspect_ratio | No | 16:9, 9:16, 1:1, 4:3 or 3:4, default 16:9, available at both resolutions. |
generate_audio | No | Off by default, so the clip comes back as picture only. Default: |
resolution | No | 720p or 1080p, default 720p. It sets the per-second rate and how long the task takes, and both resolutions accept all five aspect ratios. |
duration_seconds | No | Any whole number from 2 to 10, default 5. It is the main driver of the price, and also where likeness drifts: check the last second of a long clip before committing to that length. |
shot_type | No | single keeps one continuous take; multi lets the model cut between shots within the same scene, which suits a scene with a beat to it. |
enable_prompt_expansion | No | Off by default. Turning it on lets a one-line idea grow into a fuller brief before generation; leave it off when your prompt already names the cast and the shot precisely. Default: |
seed | No | 0 to 2147483647. Hold it while swapping one reference at a time so the random draw stays constant and the thing you changed is the only thing moving. |
Pricing
Reference clips add to the billed seconds, counted up to five seconds in total; still images are free. When a clip's length cannot be read before the job runs, we hold funds for the 5-second maximum and release the unused portion the moment the job settles, so you are only charged for the seconds actually used.
| Resolution | Generate audio: No | Generate audio: Yes |
|---|---|---|
| $0.0225 / second | $0.045 / second |
| $0.03375 / second | $0.0675 / second |
Use cases
Put two people in one scene
A portrait of each person, then a prompt for where they are and what passes between them. The rooftop example on this page starts from two stills and nothing else, with sound switched on so the scene carries its own wind and city noise.
Move a subject into a new location
A short clip or a still of the person, then the new set, the light and the camera move. The look comes from the reference while the location is written, so one shoot supplies footage for places that were never available on the day.
Combine a place with a person
A clip of a location and a still of a person cast them together: the quay example on this page uses a five-second harbour clip as character1 and a portrait as character2, and the woman walks through a place she was never filmed in.
Draft casting ideas quickly
Flash favours turnaround and cost over the last degree of texture, which suits testing several casts, settings and camera moves at 720p before you commit to a final take.
Prompt guide & pro tips
Name every reference before the action
Start with the cast — character1, an adult woman in a green slip dress; character2, an adult man in a white shirt — then the setting, the action, the camera and the light, within 1,500 characters. Re-describing what the references already show spends that budget and can compete with them. Use negative_prompt for what keeps turning up and should not: extra people, captions, or a prop that follows a subject into the new scene.
Pick references the model can read
Stills and clips where the subject is large in frame and well lit transfer best; a subject that stays small or backlit tends to come back only approximately. Stills carry colour and look well, while shape and motion come through more faithfully from a clip. Send images as JPEG, PNG without transparency, BMP or WEBP, 240 to 8,000 pixels on each side and up to 20 MB, and clips as MP4 or MOV.
Budget the seconds before you submit
Billable seconds are the output duration plus the reference clips, each rounded up to whole seconds and at least one second per clip, with the clips counted up to five seconds in total; still images add nothing. A 5-second render from two stills is billed as 5 seconds; with one 3-second clip it is 8. The rate follows the resolution and doubles with generate_audio on. The clips you upload may add up to five seconds, and the live estimate before submitting shows the amount; when a clip's length cannot be read in advance, the estimate assumes five seconds and the final charge uses the measured length.
Adapt the included example
Start from the rooftop example: two stills in reference_image_urls, aspect_ratio 16:9, duration_seconds 5, resolution 720p and generate_audio on. Replace the cast and the staging while keeping that input contract intact.
character1, an adult woman in a dark green silk slip dress, and character2, an adult man in an open-collar white linen shirt, slow-dance on a city rooftop at night, foreheads almost touching. Her hand slides up the back of his neck as he pulls her closer by the waist; she laughs quietly against his cheek. String lights sway in the wind, the city glows behind them. Slow handheld orbit around the pair, warm practical light, shallow depth of field, film grain. Sound: soft wind, distant traffic, her low laugh, a saxophone drifting up from a bar below.
Check the result before iterating
Play the whole clip rather than the first second: likeness tends to drift towards the end of the longer durations. multi mostly pays off from about 6 seconds upward, and choosing it also switches enable_prompt_expansion on for that request. Hold the seed and swap one reference at a time when comparing; it keeps the random draw constant without promising a frame-identical rerun. Generated clips are deleted after 14 days, so copy anything you want to keep.
API quick start
# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
-d '{
"model": "alibaba/wan-2.6-flash/reference-to-video",
"input": {
"aspect_ratio": "16:9",
"duration_seconds": 5,
"generate_audio": false,
"prompt": "A slow aerial shot drifting over a misty pine forest at sunrise",
"resolution": "720p",
"shot_type": "single",
"reference_video_urls": [
"https://example.com/reference.mp4"
]
}
}'
# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
-H "Authorization: Bearer $SPICY_API_KEY"
See the API reference for authentication, the complete schema, polling, and error handling.
