How to use Seedance 2.5 API via Ofox?

Create a generation task via POST https://api.ofox.ai/v1/videos with your Ofox API key, then poll the task status until it completes and download the video from mirror_urls (or unsigned_urls).

What should I do if a generation fails?

Keep the task ID and check its status and error first, then decide whether to adjust parameters or create a new task. Don't resubmit while the outcome is unclear; whether failed runs are billed follows the service rules and your bill.

Are first/last frames the same as reference images?

No. First/last frames define how the video starts or ends and use frame_images. Ordinary references guide subject, style and so on and use input_references. The two can't appear in the same request.

Why doesn't the sample change when I edit the prompt?

The workspace samples are past outputs of their original prompts, there to help you judge quality. Your edits flow into the API request; clicking “Generate video” creates the new task in OfoxAI Studio.

Is Seedance 2.5 good for videos with sound?

Seedance 2.5 supports synchronized audio, and several samples on this page include native audio you can listen to. Lip sync, pronunciation, ambient sound and character continuity still need checking against your own material — test with short clips before going longer.

How should I compare Seedance 2.5 with other video models?

Use the same input, duration, resolution and acceptance criteria, and record results, cost and failures. The samples here only show Seedance 2.5 and can't support a cross-model ranking.

Can I edit or extend an existing video directly?

A video reference isn't the same as editing or extending. The instructions and endpoints for editing or extension need to be verified separately; the workspace doesn't offer unverified edit or extend switches.

Why might advertised reference limits differ from the API's?

A model's maximum capability isn't what every provider, gateway endpoint and web form supports. Go by the limits of the endpoint you call; with many mixed assets, validate with a minimal sample first and scale up gradually.

Do the test budgets include free credits?

No. They estimate usage cost for the given settings, assume no free credits, and aren't a minimum top-up. Check your account for any credits you actually have.

Seedance 2.5

Video Generation

bytedance/seedance-2.5

Start from one shot, and set the idea in motion.

Create short clips from text, images or video references. See real samples first, pick your settings and budget, then plug the same task into your product.

4–30sper clip
480p–1080presolution
Synced audiooptional

Try a clip View API request

Billed per second generated · see this run's budget next to the settings

Seedance 2.5 is ByteDance's next-generation multimodal video creation model from the Doubao line, released on 2026-08-07. It accepts mixed text, image, video, and audio input — up to 30 images, 10 videos, and 10 audio clips, for 50 reference assets in total — and covers text-to-video, image-to-video, and video-to-video generation. Clips can run any whole number of seconds from 4 to 30, at 480p, 720p, or 1080p resolution (720p by default), with six aspect ratios plus an adaptive option. Output carries synchronized audio by default, and prompts can be written in multiple languages. This generation improves instruction following, multi-shot narrative, and long-range temporal consistency. Accessible via the OpenAI-compatible protocol.

Prompt

867 / 4000

15s × $0.24/s · Generates in OfoxAI Studio (sign-in required) · Pricing basis & multi-run budget ↓

One take through an entire cruise ship

Down the deck slide into the pool, through the dining room, kitchen and engine room, out over the wake — does an uncut long take stay coherent?

Seedance 2.5720p · 15s · 16:9 · no audioReal output

# 1. Create the generation task — keep the returned idcurl https://api.ofox.ai/v1/videos \  -H "Authorization: Bearer $OFOX_API_KEY" \  -H "Content-Type: application/json" \  -d '{  "model": "bytedance/seedance-2.5",  "prompt": "Continuous single take, camera tilting, rolling, spinning, flying through a colossal luxury cruise ship. Starts flying over the sunny upper deck dodging beach balls and sunbathers, camera dives straight down a twisting, translucent water slide, gliding past screaming children before splashing into the main pool. Sweeps underwater, effortlessly passing through a glass floor directly into the elegant grand dining room below, weaving through crystal chandeliers, clinking champagne glasses, and waiters balancing silver trays. Camera dives past the spinning kitchen doors into the humming, immaculate engine control room, transitions out the stern exhaust vent, gliding over the churning, foamy white wake as the sun sets over the ocean. No cuts, impossible camera moves, seamless transitions, energetic, dynamic, cinematic. 8K high definition, high quality footage.",  "duration": 15,  "resolution": "720p",  "aspect_ratio": "16:9",  "generate_audio": false,  "provider": {    "type": "byteplus"  }}' # Or: "provider": { "type": "volcengine" } # 2. Poll with that id until status is completed / failed / cancelled / expiredcurl https://api.ofox.ai/v1/videos/$TASK_ID \  -H "Authorization: Bearer $OFOX_API_KEY" # 3. Read mirror_urls[0] (persistent), or unsigned_urls[0] (expires in ~24h)

NextGET /v1/videos/{id} → mirror_urls[0]Full integration flow ↓

Switch sampleClick a sample to load its original prompt and settings · Sample sources ↗

Understand the model before picking an entry point

What can Seedance 2.5 do?

Seedance 2.5 is ByteDance's video generation model. It turns scene descriptions and reference media into moving shots, with synchronized audio; access it on Ofox through the unified video task API.

View model spec source

Decide what you need to control

With no assets, describe the action and camera in words. With a subject image, first decide whether you need to lock the opening frame or only borrow the subject and style. With a video, spell out the motion you want to reference.

Text → video
Best for validating a visual direction first. Describe the subject, action, scene and camera — no reference assets needed.
Check first:Is the subject clear and the action complete?
Image → video
The first frame sets where the shot starts; the last frame constrains where it ends. Ordinary reference images instead guide subject and style — a different role.
Check first:Start and end frames, subject detail, the transition.
Video → reference generation
Best for conveying motion, framing or pacing intent. Say what to reference and what to change — not just “make it like this”.
Check first:Is the motion right, and is the reference misused?

Duration: 4–30s
Any whole number of seconds, up to 30s per clip
Output resolution: 1080p
Up to 1080p, 720p by default
Aspect ratio: 7aspect ratios
21:916:94:31:13:49:16Adaptive
Audio: Audio
Picture and sound in one pass, on or off
Input: 3inputs
TextFirst/last-frame imagesReference video

More reference media doesn't mean more control. State what each asset is for: who the subject is, where the scene is, which motion must be kept. Don't let several conflicting descriptions fight over the same shot.

Released 2026-08-07bytedance/seedance-2.5

Model capability, provider support and API constraints are three different layers. Reference counts, media sizes and available combinations depend on the endpoint you use; the workspace only demonstrates a single video reference or first/last frames and doesn't treat the model's advertised maximums as form capabilities.

Results you can see

Not just the clip — how it was shot

6 different tasks, 6 original prompts. Look at motion, materials and people, then decide whether it fits your project.

Explore more prompt examples

Camera move · one take

Five spaces, no cuts

Deck, slide, pool, dining room, engine room. Watch whether the transitions flow and the spatial logic holds.

Product · vertical ad

A dessert ad from two short paragraphs

The prompt is two short paragraphs. Judge the lighting, layering and clean-background finish for e-commerce use.

Character · native audio

Write shots on a timeline — sound included

Each of the three shots specifies picture and audio. Check character consistency and how the music matches the fingers.

Animation · performance

Get two characters to interact

A head turn, a nose touch, a hug, with a push from medium shot to close-up. Check the action flow and expressions.

Timelapse · continuous change

Slow deformation has to stay stable

Petals open one by one against a clean gradient. Slow it down and look for skipped frames or flicker.

Multi-shot · cut to the beat

Ten shots, one party, fifteen seconds

Each second has its shot size and camera move, with native audio. Check pacing, look and transitions.

Where the samples come from ›

OfoxAI generated these samples with bytedance/seedance-2.5; the prompts are the exact text used. Each clip is labeled with its resolution, duration, aspect ratio and whether it has audio, and “Reuse this shot” loads those same settings. They are for judging quality only — they don't indicate success rates or support comparisons with other models.

Prompts should direct the shot

Don't just write “cinematic” — say what happens in the shot

Express subject, action, camera and constraints separately. The original templates below are drafting tools; actual results need to be verified by generating.

01
Subject and scene
Who, where, and which recognizable details must be kept.
02
Main action
How the action starts, how it changes and where it ends.
03
Camera and framing
Shot size, direction of movement, orientation, whether cuts are allowed.
04
Sound and constraints
Ambient sound or no audio; text, logos or actions that must not appear.

Choose a starting point

Show one product detail clearly

One shot, one selling point: surface finish, structure or a use gesture. Get the subject right before trying complex camera work.

Draft promptProduct still · 5s landscape

An unbranded ceramic mug on a clean wooden table. The camera pushes in slowly from a medium shot toward the rim while soft side light glides across the glaze. The mug stays still, the background is simple, and the whole clip is one continuous push-in. 5 seconds, 16:9, no captions, no extra text, no audio.

Prepare:: An accurate product description; with a product photo, use it as the first frame.
Accept if:: Outline, markings and structure don't warp.

Grab attention with one action

Build the opening and ending around one clear action. Leave room for captions in post — don't ask the model to render complex text.

Draft promptSubject action · 5s vertical

An adult in a plain tracksuit turns around once in a bright studio and ends facing the camera, standing still. The camera holds a fixed full-body frame with the subject centered throughout; clothing and limbs move naturally. 5 seconds, 9:16 vertical, single shot, no captions, text or logos, no audio.

Prepare:: Subject, action and a vertical composition.
Accept if:: The action reads clearly and the subject stays in frame.

Move the story with a reaction

Write the character's situation, action and emotional turn first. Don't hide unclear relationships behind lots of cuts.

Draft promptEmotional shift · 10s landscape

An adult sits by a quiet window, at first looking down at an envelope on the table, then slowly raising their head to look outside, their expression shifting from hesitation to calm. The camera pushes in slowly from a medium shot and settles on a close-up of the face; costume and setting stay unchanged. 10 seconds, 16:9, single shot, no captions, no extra text, no audio.

Prepare:: Character relationships, costumes and shot order.
Accept if:: Expressions, costumes and depth stay continuous.

Make a process easy to follow

Break the process into observable steps; don't make hands, tools and objects do too much at once. Leave labels for post.

Draft promptProcess demo · 8s landscape

Top-down view of a clean wooden workbench. A pair of hands folds a square sheet of paper in half, then in half again, and finishes a simple paper boat, placing it in the center of the table. The camera is locked off; the hand movements are clear and evenly paced, and the creases and deformation of the paper look real. 8 seconds, 16:9, single shot, no captions, no extra text, no audio.

Prepare:: Step order, contact points and the finished state.
Accept if:: The process flows and contact and deformation look believable.

Replaces the prompt, mode and settings in the Playground. Nothing is generated automatically.

Change one variable at a time

Round one: composition. Round two: the action. Round three: lighting or pacing. Keeping each round's prompt and result makes problems far easier to pin down than rewriting everything at once.

Know the cost up front

What does one clip cost — and three tries?

The budget follows the current settings. Validate the direction with short clips first, then decide whether to add duration or resolution.

Test budget for the current setup

Text to video · 720p · 15s

One clip

$3.60

3 runs total

$10.80

Plan 3 rounds

Estimated cost$10.80

Not a minimum top-up. Test runs are not usable clips; actual charges follow your live bill.

Billed per generated second
Resolution	Text to Video / s	Video to Video / s
480p	$0.11/s	$0.14/s
720p	$0.24/s	$0.3/s
1080p	$0.6/s	$0.71/s

Live Ofox prices — the same rates used by the budget on the left. Image-to-video is not priced separately here, so no rate is assumed. Listed tiers apply to any audio setting. Several platforms serve this model; set provider.type in the request to pick one.

View live price source

Split your first validation into three stages

Suggested text-to-video test setups; costs are calculated live from current prices.

01 / Direction4s · 480p$0.44/ run
Confirm the subject and action are understood at the lowest cost; if not, fix the prompt and retry before upgrading.
02 / Detail5s · 720p$1.20/ run
Move to the delivery setup only once structure, edges, contact and continuity hold up; fix issues at this tier first.
03 / Delivery5s · 1080p$3.00/ run
Use only after direction and detail are settled, to confirm the final look at the target resolution.

From trying to building

Happy with the shot? Ship the same settings in your app

Create a video task with the current prompt and parameters — no need to reassemble the request from the docs.

1POST /v1/videos
Create a video task
Send the model, prompt, duration, resolution and audio setting. Reference media goes into the request according to the mode.
2GET /v1/videos/{id}
Query the status
Keep the returned task ID. Query until completed, failed, cancelled or expired; don't create the same task twice.
3mirror_urls / unsigned_urls
Get the video result
Check the task status first, then read the result URL. Prefer mirror_urls; fall back to unsigned_urls when missing.

Endpoints and fields follow the official code examples.

Read the developer docs

What problem does each field solve?

Full field reference and request examples

What you want to expressOfox fieldWhat to watch for

Which model: model; Fixed to Seedance 2.5 on this page; don't use other platforms' routing IDs.
What happens: prompt; Describe subject, action, camera and end state; reference tasks need a description too.
Duration and framing: duration / resolution / aspect_ratio; Manual duration on this page is 4–30s; first/last-frame tasks use the adaptive aspect ratio.
Pin the start or end: frame_images; Use first_frame / last_frame to set each frame's role; don't mix with ordinary reference fields.
Guide content or style: input_references; Image, audio and video references aren't exact frame anchors; constraints follow the current endpoint.
Generate sound or not: generate_audio; Controlled separately from visual quality. Samples with audio off don't mean the model lacks it.
Pin the hosting platform: provider.type; When several platforms host the same model, requests are split by weight unless you set this.

Where do first integrations get stuck?

Tell “not accepted” apart from “accepted but not finished”. Once creation returns a task ID, query that task first.

Parameters or media don't match

Check duration, enums, URLs and input type. First/last frames and ordinary references use different fields and can't be mixed in one request.

Account or balance issues

Verify your server-side key and available balance. Resolve authentication and billing first, then try creating again.

The task is still waiting

Keep the ID and query until a definite terminal status, with backoff to avoid frequent polling. A front-end timeout doesn't mean generation failed on the server.

Results need long-term storage

Store the video in your own delivery storage once you have it. Don't treat temporary result URLs as a permanent library; confirm the task completed before downloading.

Already using an agent workflow?

Hand over the whole task, not just “make me a video”

Provide the goal, duration, aspect ratio, what each asset is for, a budget cap and acceptance criteria. Have the agent show the request and estimated cost first, create the task only after you confirm, keep the task ID and return the result or a clear error.

This is a copyable task brief; it doesn't connect to any agent or run any call. Keys stay in your server-side environment.

Questions before you start

Seedance 2.5 FAQ

Seedance 2.5 on Ofox.ai costs from $0.11 per second of generated video. The exact rate depends on resolution, generation mode, and audio. Pay-as-you-go, no monthly fees.