Introduction
Send a prompt plus at least one of reference_image_urls, reference_video_urls or reference_audio_urls. Each reference keeps its own job in the scene: identity, product, camera move or voice. Duration, resolution, aspect ratio and audio are optional.
| Field | Value |
|---|---|
| Input | Text |
| Output | Video |
| Publisher | ByteDance (Seedance 2.5 Flash, speed tier of Seedance 2.5, announced by BytePlus on 12 November 2026) |
| API names | bytedance/seedance-2.5-flash/text-to-video, image-to-video, reference-to-video, video-edit, video-extend |
| Resolution | 480p or 720p (default 720p); 1080p and 4K not supported |
| Clip length | 4–30 whole seconds, default 5; video-edit follows its 4–30 s source |
| Aspect ratio | 16:9 (default), 9:16, 4:3, 3:4, 1:1, 21:9 or adaptive on text and reference requests; image-to-video follows the start image |
| Reference limits | Up to 10 images, 5 clips (30 s in total) and 5 audio files on reference-to-video |
| Reference clip rules | Each clip counts as at least 2 s and needs about 409,600 pixels (854×480 or larger) |
| Edit inputs |
|
| Audio |
|
| Output options | mp4 (default) or mov; watermark off by default; |
| Billing basis | Per second by resolution; reference clips add their seconds; edits bill source seconds; failed generations are free |
| Measured render time | 84 s to about 11 min for 4 s at 480p (13 November 2026 runs) |
Key features & highlights
Flagship workflows at a lower ceiling
Flash keeps the four Seedance 2.5 workflows that fit drafting (text, start and end frames, mixed references, clip editing), drops 1080p and 4K, and adds a video-extend built on its image-to-video mode. BytePlus presents the trade as results close to the flagship at nearly twice the speed. Our 4-second 480p renders took anywhere from 84 seconds to about 11 minutes, so measure speed on your own queue.
Give every upload a job in the prompt
Name uploads by type and order with a space: @Image 1, @Video 1, @Audio 1. With two portraits as @Image 1 and @Image 2 in 16:9, both people appeared together in one elevator scene and kept their faces, hair and clothes. With one image, one clip and one voice file, the clip's street, yellow tram and tracking camera carried over, but the person stayed back to the camera and the face of @Image 1 barely showed.
A sunny street turned into a snowy evening
Both riders, their bikes, the tram and the camera move matched the source frame by frame, and the weather and light changed while nothing else did. That was a 4.0 s clip at 480p with audio off, which also kept the original soundtrack.
One switch decides whose soundtrack you get
generate_audio is on by default and produces speech, effects and ambience with the picture. In video-edit, switching it off keeps the source clip's own track (44.1 kHz in our test), while switching it on generates a new one that can follow @Audio 1. The switch has no effect on the price.
Request parameters
| Parameter | Required | Description |
|---|---|---|
prompt | Yes | Name each upload with a space and its type number, such as @Image 1, @Image 2, @Video 1 or @Audio 1, and give it one job: identity, product, camera move or voice. at least 1 characters · at most 20000 characters |
reference_image_urls | No | Up to 10 images of people, products, places or a look to keep. Send one clear, well-lit picture per subject and call it out by name in the prompt. at least 1 items · at most 10 items |
reference_video_urls | No | Up to 5 clips, 30 seconds in total, each at least about 409,600 pixels: 854×480 and 1280×720 pass, while a 640×360 clip was refused at submit. at least 1 items · at most 5 items |
reference_audio_urls | No | Up to 5 voice, music or sound files, 30 seconds in total. Point to one in the prompt, for example "in the voice from @Audio 1". at least 1 items · at most 5 items |
aspect_ratio | No | 16:9 (default), 9:16, 4:3, 3:4, 1:1, 21:9, or adaptive to let the model choose the shape. Our two-image test with 9:16 came back at 480×854. |
resolution | No | 480p or 720p, default 720p; Flash does not offer 1080p. The resolution you pick sets the per-second rate for the attached clip seconds as well. |
duration_seconds | No | Whole seconds from 4 to 30, default 5. The requested seconds are billed in addition to the seconds of any attached clips, so a short scene with a short clip keeps drafts small. |
generate_audio | No | On by default, so the scene gets synchronized speech, effects and ambience; audio files you attach are there to guide that sound. Off returns a silent file at the same price. Default: |
output_format | No | mp4 (default) or mov. Pick the container your editor expects; the price is the same either way. |
return_last_frame | No | Off by default. On also returns the final frame as a JPEG, which you can use as the start frame of a follow-up image-to-video shot to carry the same cast forward. Default: |
watermark | No | Off by default. On adds the model's visible watermark to the frames, which you may want when a draft built from your own references will circulate; the price does not change. Default: |
Pricing
Billed per second at the 480p or 720p rate. When reference clips are attached, their seconds (at least 2 per clip) are added to the requested seconds and the whole total uses the lower with-clips rate. Reference images and audio files, audio on or off, aspect ratio, watermark, last frame and format do not change the price; failed generations are not charged.
| Resolution | With video: Yes | With video: No |
|---|---|---|
| $0.0706 / second | $0.115 / second |
| $0.147 / second | $0.2392 / second |
Use cases
Trying a shot plan before a final render
Send a 4-second, 480p, 16:9 prompt that ends with an Audio: line, such as a rain-soaked night crosswalk. Our run returned 854×480 at 24 fps, 4.064 s long, with an AAC stereo track of rain and laughter. Change one phrase and rerun to see what moved.
Restyling footage you already shot
Upload a 4–30 second clip, describe the change, then list what must stay. A sunny street with two cyclists and a tram became a snowy evening in our test, and with generate_audio switched off the original soundtrack came back unchanged.
Character and product mockups
Attach a portrait and a product photo as @Image 1 and @Image 2 and ask for 9:16. Our two-image test returned 480×854 with the man lifting a glass dropper bottle toward the lens, so both references showed up. Use the result to judge the idea before building a polished version.
Looping a portrait animation
Give the same photo as image_url and last_image_url and describe the motion between them. Our potter portrait returned a 562×748 clip of 4.064 s whose opening and closing frames both matched the source pose.
Chaining one shot into the next
Set return_last_frame to true and the MP4 arrives with a JPEG of its closing frame (854×480 at 480p 16:9 in our test). Pass that JPEG as image_url of the next image-to-video request so the following shot begins where the last one stopped.
Prompt guide & pro tips
Choose the endpoint for your input
Use text-to-video for a prompt alone, image-to-video for a start frame, reference-to-video for images, clips or a voice, video-edit for an existing 4–30 second clip, and video-extend to add 4–30 seconds to a clip of up to 60 seconds. All five share the 480p/720p range and the
generate_audioswitch.Send the createTask request
POST to https://api.spicyapi.ai/api/v1/jobs/createTask with the header Authorization: Bearer YOUR_SPICY_API_KEY and the body {"model": "bytedance/seedance-2.5-flash/text-to-video", "input": {"prompt": "A lighthouse keeper climbs a spiral stair at dawn and wipes fog from the lamp glass. Audio: wind, boots on iron.", "
duration_seconds": 4, "resolution": "480p"}}.Poll for the finished clip
The response gives a taskId in state queued. Call GET /api/v1/jobs/recordInfo?taskId=YOUR_TASK_ID, or pass a callBackUrl, until the state is succeeded, then read the MP4 address from output.assets. Field details are in the task guide at docs.spicyapi.ai/docs/tasks.
Change one phrase, then move up
Compare variations at 480p and 4 seconds, altering a single phrase per run. When a take works, rerun it at 720p and your target length, or send the same brief to Seedance 2.5 for 1080p.
What Seedance 2.5 Flash reference-to-video does
Seedance 2.5 Flash reference-to-video is ByteDance's multi-reference endpoint for the Flash tier of Seedance 2.5. You attach up to 10 images, 5 video clips and 5 audio files, and the prompt says what each contributes to a new 4 to 30 second scene at 480p or 720p, with sound.
A reference supplies identity, product look, camera behaviour or voice; none becomes the opening frame, so a shot that must start on one exact picture belongs on image-to-video. BytePlus, ByteDance's cloud platform, describes Flash as "results close to flagship Seedance 2.5, with nearly twice the generation speed", at 480p and 720p only.
Request example
Send POST https://api.spicyapi.ai/api/v1/jobs/createTask with an Authorization: Bearer YOUR_SPICY_API_KEY header and this body (task API guide). Each type is numbered on its own in list order: the portrait is @Image 1, the clip @Video 1, the voice file @Audio 1.
{
"model": "bytedance/seedance-2.5-flash/reference-to-video",
"input": {
"prompt": "The man from @Image 1 walks past the vintage car at the roadside gas station from @Video 1, filmed with the same slow camera move. He stops by the pump, looks into the lens and says a short friendly greeting in the voice from @Audio 1.",
"reference_image_urls": ["https://example.com/man-portrait.jpg"],
"reference_video_urls": ["https://example.com/gas-station-1280x720.mp4"],
"reference_audio_urls": ["https://example.com/greeting-voice.mp3"],
"duration_seconds": 4,
"resolution": "480p",
"aspect_ratio": "16:9",
"generate_audio": true
}
}
Response example
Submission returns {"code": 200, "msg": "success", "data": {"taskId": "YOUR_TASK_ID", "state": "queued"}}. Poll GET /api/v1/jobs/recordInfo?taskId=YOUR_TASK_ID or pass a callBackUrl; a finished task looks like this:
{"code": 200, "msg": "success", "data": {"taskId": "YOUR_TASK_ID", "state": "succeeded", "output": {"assets": [{"url": "https://example.com/seedance-flash-reference.mp4"}]}}}
Parameters
| Parameter | Values |
|---|---|
prompt | Required; names uploads as @Image 1, @Video 1, @Audio 1 |
reference_image_urls | Up to 10 images |
reference_video_urls | Up to 5 clips, 30 s in total, each at least about 409,600 pixels |
reference_audio_urls | Up to 5 audio files, 30 s in total |
duration_seconds | 4 to 30, default 5 |
resolution | 480p or 720p, default 720p |
aspect_ratio | 16:9 (default), 9:16, 4:3, 3:4, 1:1, 21:9, adaptive |
generate_audio, watermark, return_last_frame, output_format | true, false, false, mp4 by default (mov also) |
Send at least one image, clip or audio file; there is no image_url field here.
How billing works
Billing is per second at the rate of the chosen resolution. Without reference clips, the billed length is the seconds you request.
With clips attached, it is the requested seconds plus the seconds of every clip (each counts for at least 2) at a lower per-second rate; the price on this page lists both tiers, so trim clips to the stretch you need. Images, audio files, the audio switch, aspect ratio, watermark, last frame and format do not change the price. Failed generations are not charged.
Tips for good results
Give every reference a job; a bare tag leaves the model guessing:
| Name | Job | Wording we used |
|---|---|---|
@Image 1 | Identity | "The man from @Image 1" |
@Image 2 | Product | "the glass dropper bottle from @Image 2" |
@Video 1 | Camera move | "filmed with the same slow camera move" |
@Audio 1 | Voice | "in the voice from @Audio 1" |
Write the name with a space, as in @Image 1, the form our examples use. Numbers count per type, so the clip at the top of your video list is @Video 1 whatever the number of images.
Ask for the face. In one test, a portrait, a 4-second clip and an audio file gave the clip's street, yellow tram and tracking camera, but the person stayed back-to-camera and the face of @Image 1 barely showed. Say where the face goes, as "looks into the lens" does in the gas-station example.
Decide what the clip carries. The setting came along with the camera move in that test; to take only the move, say so.
What our tests showed
All on 13 November 2026 at 480p, 4 seconds, audio on.
| References | Result | Time |
|---|---|---|
| Portrait, 1280×720 clip of 5.04 s, voice file | 854×480, 4.064 s; the man passes the vintage car and smiles at the lens | 2.5 min |
| 2 images (man, dropper bottle), 9:16 | 480×854; he lifts the bottle and turns it to the camera, both references used | 5.4 min |
| Portrait, 4 s clip, audio file | Street, tram and camera kept; person back-to-camera | 9.4 min |
Across all four Flash endpoints, 4-second 480p clips took 84 seconds to about 11 minutes; 720p or longer clips take longer. These Flash-only timings neither confirm nor dispute BytePlus's comparison with Seedance 2.5.
Limits and errors
| Limit | What happens | What to do |
|---|---|---|
| Over 10 images, 5 clips or 5 audio files, or clips over 30 s in total | Validation fails | Keep references that carry a job; trim clips |
| A clip under about 409,600 pixels | createTask is refused at submit, no render starts; our 640×360 clip (230,400 pixels) was refused | Re-export at 854×480 or 1280×720 or larger; a vertical 480×854 clip passes too |
Use only faces, voices and footage you have the right to use, and do not present a generated person as someone real. The model runs its own copyright and likeness checks and can decline a request.
When to choose a sibling
- Flash image-to-video: one exact picture must open the shot.
- Flash text-to-video for a prompt alone; Flash video edit to change an existing clip.
- Seedance 2.5: delivery above 720p, 4K or clip extension.
- Seedance 2.0 Fast: previous-generation fast tier with extend.
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: $SPICY_IDEMPOTENCY_KEY" \
-d '{
"model": "bytedance/seedance-2.5-flash/reference-to-video",
"input": {
"aspect_ratio": "16:9",
"duration_seconds": 5,
"output_format": "mp4",
"prompt": "A slow aerial shot drifting over a misty pine forest at sunrise",
"resolution": "720p",
"reference_image_urls": [
"https://example.com/reference.png"
]
}
}'
# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
-H "Authorization: Bearer $SPICY_API_KEY"
See the API reference for authentication, the complete schema, polling, and error handling.










