Skip to main content

​
Overview

doubao-seedance-2-5-260628 (2.5), doubao-seedance-2-0-260128 (standard), doubao-seedance-2-0-fast-260128 (fast), and doubao-seedance-2-0-mini-260615 (mini/lite) are ByteDance’s latest video generation model family — four models running in parallel, served through APIYI on official Volcengine Mainland China resources (not the BytePlus international edition) with upstream content-safety built in. They support text-to-video, first+last/first frame image-to-video, and multi-modal reference-to-video — and can generate voice, sound effects, and background music synchronized with the visuals. 2.5 is the most capable tier: the duration cap goes from 15 to 30 seconds, reference images from 9 to 30, audio can stand alone as a reference, and it adds mov output plus explicit video-edit/extend task types. It also costs more — roughly 1.5x the 2.0 standard model (about $1.35 versus $0.91 for 720p/5s), matching the gap between the two generations in Volcengine’s own list prices. The 2.0 family stays available and is not being retired: for routine clips under 15 seconds the standard model is cheaper and also supports 1080p; mini is the volume-production pick (about half the standard model’s unit price and faster generation, capped at 720p), with fast in between. 2.5 runs on the same SeeDance2 group as the 2.0 family (0.18x) — one token reaches all four models. See “Group Setup” below.
🎬 Highlights: 2.5 supports 4-30 s, the 2.0 family 4-15 s (both accept -1 for a model-chosen length); three resolution tiers (480p/720p/1080p, where 1080p is limited to 2.5 and 2.0 standard); 6 aspect ratios plus adaptive; synchronized audio on by default; multilingual prompts. Built for short-video production, e-commerce assets, motion design, and virtual-human content at scale.
🔥 Limited-time discount — through 2026-10-07 23:59 (UTC+8): two new single-model groups, SD2Mini (0.10x) and SD2Fast (0.15x), cut the rate by 44.4% for mini and 16.7% for fast. Swap in one new Token to get it — no code changes. See “Limited-time discount groups” and “Group Setup” below.
Tokens calling 2.5 need the SeeDance2 group (rate 0.18x) — the same group as the 2.0 family. A token on the Default group or another video group calling doubao-seedance-2-5-260628 fails with “no available channel for this model”.One Token with SeeDance2 enabled reaches all four models — 2.5 and the 2.0 family share the same group and the same rate, so your code only changes the model field.

Video Generation API Reference

POST /seedance/api/v3/contents/generations/tasks — async task endpoint with an interactive Playground and full polling/download code.

API Manual

Token creation, base URL, billing models, and general calling conventions.

Visual API Testing

Debug this endpoint directly in the iCover visual testing tool — no code required.

Async Task Lookup / Download

View submitted video tasks and download video links in the APIYI console — a lookup entry outside the API.

​
Let an AI Agent Do the Integration

If you build with Codex / Claude Code / Cursor, copy the prompt below and hand it to your agent. It first fetches the plain-text version of this page (append .md to any docs URL), then writes code in your project’s own stack — async polling, the 24-hour link expiry that forces an immediate copy, the gzip-header trap and the parameter red lines are already baked into the requirements.

Have a coding agent integrate or troubleshoot Seedance 2.5 / 2.0 video generation. Copy and paste into Codex, Claude Code, Cursor and similar tools.

​
Why APIYI’s Seedance?

A note on positioning first: this model carries no official discount, and APIYI doesn’t price it for profit — it is offered to secure supply and serve customers. The real value of going through APIYI is not “cheaper”, but access and experience:

Official Resource · Mainland Edition

Official Volcengine Mainland China resources (not BytePlus international), with upstream content safety built in. Parameters, responses, and billing match the official API exactly.

Virtual-face Whitelist Access

The channel comes with upstream virtual-face whitelist access: AI-generated faces and virtual avatars can be used directly for image-to-video, no separate whitelist application to the official channel needed (real human faces remain restricted by upstream content safety).

Asset Library Included Free

The private asset library behind character-consistent video (virtual-avatar ingest + real-person verification) is free on APIYI. Officially it is a separately purchased add-on — a six-figure CNY annual contract for customers without a framework agreement. We include it in the API price.

Supply-first Pricing · On Par with Official

No official discount exists, and APIYI doesn’t profit from this model: unit prices align with Volcengine’s official list (on-platform billing runs roughly 10% higher); combined with top-up bonuses the effective cost is about on par with the official channel, and high-tier recharge customers can land below it on some tiers.

Unlimited Concurrency · No Queuing

In our tests, 15 simultaneous tasks all entered running immediately with zero queuing (measured 2026-06-06 (UTC+8)) — ready for batch production at scale.

Zero-friction Access · No ID Verification

No Volcengine account, no real-name/identity verification, no spending threshold (skip the CNY 200 activation deposit and enterprise verification). Mainland China data centers, residential networks, and overseas nodes can all reach api.apiyi.com directly with a single Token.

Full Video Model Lineup

VEO 3.1 and Wan2.7 are available on the same platform — mix and match per use case.

Professional Support

A team experienced in video-generation workloads, providing model selection, tuning, and integration support from PoC to production.

​
Key Features

Three Tiers · Same Price per Tier

480p / 720p / 1080p (1080p on 2.5 and 2.0 standard only; fast and mini cap at 720p). Within a tier, 16:9, 9:16, 1:1, and every other ratio share the same pixel area and the same price — switch between landscape and portrait at zero cost.

Synchronized Audio by Default

generate_audio defaults to true: voice, sound effects, and background music are generated to match the visuals. Put spoken lines in double quotes to improve voice-over quality.

Up to 30 s Controllable Duration

2.5 accepts whole seconds from 4 to 30, the 2.0 family 4 to 15; -1 lets the model pick a length (billed by actual output). On 2.5 the duration default is -1 — omit it and the model chooses for you. Fixed 24 fps.

Multilingual Prompts

Chinese (up to ~500 chars) and English (up to ~1000 words), plus Japanese, Spanish, Portuguese, and Indonesian.

First+Last / First Frame

Pin both first and last frames with two images, or animate a single image as the first frame. Combine with return_last_frame to chain clips into longer continuous videos.

Multi-modal Reference-to-Video

2.5 takes up to 30 images + 10 videos + 10 audios, and audio can be used on its own; the 2.0 family takes 9 images + 3 videos + 3 audios, with audio requiring an image or video alongside it. Create, edit, or extend videos while keeping characters and style consistent.

Async Task Flow

Submit and get a task_id, poll for status, then download the mp4 from content.video_url (link valid for 24 hours).

Reproducible Seeds

Fix seed for similar results across runs. watermark defaults to false — output is watermark-free.

​
Pricing

Pricing in one line — precise token billing, tier-by-tier pegged to the Volcengine official site. The four models are priced differently: mini < fast < standard < 2.5 (same direction as the official site; mini runs at about half the standard model’s unit price and 2.5 at about 1.5× it — they are NOT the same price level). Volcengine offers no discount on this series, and this channel is priced to secure supply, so the nominal rate sits slightly above the official one; with the top-up bonus (10% for general customers, up to 20% for large-deposit customers) the effective cost is essentially on par with the official site — large-deposit customers pay only about 5% more, and at some tiers (e.g. 1080p large-customer price) even less. Billing is by area×duration, so a ±5% deviation is normal — you’re welcome to test, reconcile, and reach out anytime.One more thing: the asset library required for character-consistent video is included free on APIYI (officially a separately purchased add-on — a six-figure CNY annual contract for customers without a framework agreement). That value is not reflected in the unit-price comparison above.
Token-based billing: tokens ≈ (input video duration + output duration)(s) × output width × output height × 24 / 1024 (input video duration is 0 for text-/image-to-video; verified in our tests to within 0.1%). Since every ratio in a tier has the same pixel area, price depends only on the resolution tier, output duration, and whether the input includes video.

​
Official price anchors (16:9 / 5 s output, CNY per video)

① No input video (text-to-video / image-to-video / reference images): ② With input video (multi-modal reference incl. video_url; input video 2-15 s, low end ≈ 2-4 s input, high end ≈ 15 s input):
With input video, billed duration = input video duration + output duration, so it costs more than plain text-/image-to-video; a minimum-token floor also applies (very short inputs bill at the floor). The authoritative usage is the returned usage.completion_tokens.
Measured on-platform comparison (tested 2026-06 and 2026-07, 16:9 / default audio / no input video; CNY at the fixed 1:7 rate, for reference only):
The three models are NOT the same price — never treat them as equal. At the same resolution/duration, per-token price goes mini < fast < standard (e.g. 720p/5s: mini ≈ ¥3.16, fast ≈ ¥5.08, standard ≈ ¥6.35), matching the official price ladder. For batch production, mini saves the most money and time (our 2026-07 tests measured its effective unit price at exactly the platform’s nominal rate, 0.00% deviation); 1080p is standard-only.
Note: “CNY” is the on-platform list charge; “General ÷1.1” and “Large customer ÷1.2” are the effective prices after a 10% / 20% top-up bonus — after the bonus, prices land close to the official reference, and the 1080p large-customer price is even below official. The authoritative usage is the returned usage.completion_tokens.

​
Seedance 2.5 Pricing (SeeDance2 group, 0.18x)

2.5 and the 2.0 family share the same SeeDance2 group and the same 0.18x rate — the gap between generations comes entirely from the models’ own unit prices. 720p/5s costs $1.3721 on 2.5 versus $0.9074 on the 2.0 standard model, roughly 1.5x. That gap mirrors Volcengine’s own list prices (their 2.5 token rate is about 52% above 2.0); it is not an APIYI markup. Whether the upgrade is worth it comes down to whether you actually need 30-second clips, 30 reference images, mov output, or video editing/extension — if you do not, the 2.0 standard model is cheaper and also supports 1080p. ① No video in the input (text-to-video, image-to-video, reference images): ② Video in the input (multi-modal reference with video_url, video editing, video extension): billed at a separate, lower token rate. A lower rate does not mean a lower bill: with video in the input, billable tokens = (input video duration + output duration) × area, so the token count itself grows. The row above would cost $1.0945 at rate ①; on rate ② it is $0.6567. The two token rates: Both the token counts and the two rates come from actual billing logs (re-verified 2026-08-31), not from a rate calculation. After stacking the recharge bonus, high-tier recharge customers land close to the official reference price.
The two easiest ways to overspend on 2.5:
  1. duration defaults to -1 (the 2.0 family defaults to 5). Omit it and 2.5 picks its own length between 4 and 30 seconds — a test request with no duration returned a 10-second clip, exactly twice the cost of 5 seconds. Pass duration explicitly when cost matters.
  2. 30 seconds costs 6× what 5 seconds costs (¥57.23 vs ¥9.60 at 720p). Cost scales strictly with duration, so validate the prompt at 5 seconds before committing to long clips.
The tables above and the 2.0 table further up are both on the SeeDance2 group at 0.18x, so they can be compared directly. fast and mini also have limited-time discount groups with lower prices — see the next section.
Billing notes:
  • Final charges follow the console’s model pricing and call logs
  • Tasks are pre-charged on submission and settled on completion — your balance fluctuates briefly; reconcile against call logs, where one video produces two charge entries (see “Reading charges in the logs” below)
  • The hold is based on duration alone, independent of resolution: $0.09/second for the 2.0 family, $0.135/second for 2.5. That is why 1080p usually settles with an extra charge and 480p usually gets a partial refund — both are normal
  • Rejected requests (HTTP 400 parameter errors, etc.) are not billed (verified)
  • Cost scales linearly with duration: a 15 s video costs about 3× a 5 s one

​
Limited-time discount groups (mini / fast only, through 10/7)

Two limited-time discount groups launched 2026-08-08: SD2Mini (0.10x rate) and SD2Fast (0.15x rate). Against the regular SeeDance2 group at 0.18x, that is 44.4% off for mini and 16.7% off for fast. Model capabilities, parameters, endpoints, and call syntax are unchanged — swap in one Token, leave your code alone. Offer runs through 2026-10-07 23:59 (UTC+8) (extended on 2026-09-05 in step with Volcengine’s official promotion; the original end date was September 7).
Like-for-like price comparison (scaled from the measured figures above in proportion to the new group rate, that is, divided by 0.18 and multiplied by the new rate; CNY converted at the fixed 1:7 rate, for reference only): Same model, same spec: the discount groups save 44.4% on mini and 16.7% on fast. This is a straight group-rate reduction and stacks independently with the top-up bonus. Billing is unchanged — still settled on actual token usage, with ±5% deviation being normal.
Nothing gets cut off when the offer ends: after 2026-10-07 23:59 (UTC+8), SD2Mini and SD2Fast stay online — their rate simply reverts to 0.18x, matching the regular SeeDance2 group. Your Tokens keep working and no code changes are needed. If you have batch production planned, schedule it inside the discount window.

​
Reading charges in the logs (pre-charge + settlement)

Open the console log page at api.apiyi.com/log and search for the model name doubao-seedance-2-0 to see every charge. One video produces two charge entries:
  1. Pre-charge: an estimated amount deducted when the task is submitted (log entry labeled “non-streaming”, showing the token and group) — $0.449998 in the screenshot below
  2. Settlement (charge or refund): after the task completes, the difference is settled against the actual generated tokens (log entry labeled “streaming”, with a completion-token count) — $5.611858 below; 1080p usually incurs an additional charge
APIYI log page showing the two charge entries for one Seedance 2.0 video: pre-charge and settlement

Two charge entries for one 15 s 1080p video: pre-charge + settlement

The settlement entry shows neither the token nor its group — this is normal. The sum of the two entries is the video’s total cost.
How to read the time fields:
  1. The first entry’s (pre-charge) timestamp is the video’s submission time; its “first byte” value is how long the submission took to return a task ID (e.g. 首字节:3秒 / first byte: 3 s) — not the generation time
  2. The settlement entry shows 流式 (streaming) and 首字节:<1秒 (first byte under 1 s) — these are just internal markers on the settlement record, not a sign of any problem
  3. The video’s actual generation time is the “耗时” (elapsed) column on the “Async tasks” page (api.apiyi.com/task) in the top navigation
Reading the time and first-byte fields on the log page: the first entry is the submission time and submission latency

The first log entry's timestamp = submission time, and its first-byte value (3 s) is the submission latency; this fast example settled as a refund (negative amount), total cost 0.360000 − 0.022750 = 0.337250 USD

The Async tasks page showing each video task's submission time and generation elapsed time

The elapsed column on the Async tasks page is the actual video generation time, e.g. 158 s, 303 s

For the 15 s 1080p video in the first screenshot, the total cost = 0.449998 + 5.611858 = $6.061856. The matching task parameters are visible under “Async tasks” at the top of api.apiyi.com/task, and they line up exactly with the charges:
The 732,108 completion tokens ≈ 15 × 1248 × 1664 × 24 / 1024 (a 3:4 video at 1080p outputs 1248×1664) — consistent with the billing formula.
This 15 s 1080p video totals about ¥42.4 (nominal charge at the fixed 1:7 rate); with the top-up bonus the effective cost is roughly ¥35-39, versus an official reference of about ¥37.2 for the same spec. The official pricing itself is not cheap — cost is driven by model + resolution + duration (switching to fast / 720p / 5 s is far cheaper). This model is supplied at a thin margin to secure availability, and large-deposit customers get bigger discounts.
Beta-supply notice: Seedance 2.5 and the 2.0 family are currently in a beta supply phase. If your actual charges deviate noticeably from the table above, contact customer support and we will reconcile. Pricing will be adjusted dynamically with upstream policy (e.g. if a lower-priced official variant ships later) and APIYI’s supply capacity; capable channel partners are welcome to reach out. This model is priced to secure supply and serve customers, not for profit.

​
Group Setup

Seedance 2.5 and the 2.0 family run on a dedicated group, with two hard requirements: ① the Token’s billing model must be Pay-as-you-go Priority (or Pay-as-you-go) — Pay-per-request tokens cannot route; ② the Token must have the matching group enabled. Tokens on the Default group or other video groups will fail with “no available channel for this model”. There are three groups today. 2.5 and the 2.0 family share SeeDance2, plus two limited-time discount groups that each serve exactly one model:
One SeeDance2 token reaches all four models: 2.5 and the three 2.0-family models all live in this group, so your code only changes the model field.The two discount groups are single-model channels: SD2Mini carries only mini and SD2Fast only fast, so calling any other model through them returns the same error.No cutoff when the offer ends: after 2026-10-07 23:59 (UTC+8) both discount groups stay online with the rate reverting to 0.18x — no Token or code changes required.
2.5 uses the same 0.18x. The gap between generations comes from the models’ own unit prices (Volcengine prices 2.5 above 2.0), not from the group rate — so 2.5 and the 2.0 family run on exactly the same conversion basis.Why 0.18x? The system’s built-in unit prices for Seedance match Volcengine’s official list prices — but that list is denominated in CNY, while APIYI balances are denominated in USD (fixed 1:7 USD/CNY rate). A 1x rate would effectively charge 7× the official number, so the group rate is lowered to absorb the currency conversion — that is where 0.18 comes from: not a discount, and not a markup.Volcengine offers no discount on this series, and this channel is priced to secure supply. The nominal charge sits slightly above the official reference price by default, and the top-up bonus largely absorbs it: large-deposit customers pay only about 5% more, and some tiers (1080p, for instance) land below the official price.Please be aware: billing is always based on actual token usage, and token conversion carries a small natural variance (±5% is normal); the official list price is only a reference anchor, not a per-request guarantee. The current pricing is a reasonable supply-first arrangement — always evaluate it together with the recharge bonus. If a charge looks off, we’re happy to reconcile bills with you anytime; however, “why is it slightly above the official price” is not up for debate — please keep this in mind, and skip this channel if that is a concern. On the flip side, ample concurrency with no queuing is exactly what this channel delivers.On top of that, the asset library (virtual-avatar ingest / real-person verification) is included free on this channel — officially it takes a six-figure CNY annual contract to purchase separately. That is part of what this channel is actually worth.

​
How to set up your Tokens

If you are not chasing the discounts: create one Token with the SeeDance2 group enabled — it reaches all four models — and skip the table below. If you want the limited-time discounts: mini and fast have single-model groups of their own, so split Tokens as below: Every Token must use the Pay-as-you-go Priority (or Pay-as-you-go) billing model. If you only use mini, one Token A is enough; if you only use 2.5, one Token C is enough — you do not need all of them.
Why a separate discount Token is worth it:
  • You won’t miss the deadline — billing is split per Token, so what you used and saved during the offer is visible at a glance, and it is easier to decide whether to pull batch work forward as October 7 approaches
  • Switching costs nothing — when the offer ends, point your client back at Token C; no code changes, no group edits
  • Dedicated Tokens are the production recommendation anyway — per-business-line quota control and alerts, and far easier to trace an unexpected spike
Seedance 2.5 is live (2026-08-28): the model name is doubao-seedance-2-5-260628, on the same SeeDance2 group (0.18x) as the 2.0 family. Endpoint, auth, and request shape are identical to 2.0 — swap the model field and your code keeps working. Against the 2.0 family: duration cap 15 s → 30 s, reference images 9 → 30, reference videos/audio 3 → 10, audio usable on its own, plus mov output and the omni_reference_task_type task selector. It runs at roughly 1.5× the 2.0 standard model. See “Technical Specs” below for the full diff.

​
Technical Specs

​
API Endpoints

Domains: api.apiyi.com is the primary gateway; vip.apiyi.com and other platform domains behave identically. The path prefix is /seedance/api/v3 — do not drop the /api segment, and do not use /v1/videos.

​
Resolutions & Aspect Ratios in Detail

A resolution tier defines the pixel area, not the short side. Actual output dimensions per ratio (official values, verified in our tests):
On 2.5, 480p 16:9 is 854×480 (measured), not the 864×496 that 2.0 produces — slightly smaller in area, so the same spec costs slightly fewer tokens. Every other tier we measured matches across both generations: 480p 1:1 = 640×640, 720p 16:9 = 1280×720, 720p 21:9 = 1470×630, 1080p 16:9 = 1920×1080.No model in this family supports 4k — sending "resolution": "4k" returns a synchronous 400 (not billed).

​
How adaptive works

  1. Text-to-video: the model infers the best ratio from your prompt
  2. First+last / first frame: matches the first-frame image’s ratio (mismatched images are center-cropped)
  3. Multi-modal reference-to-video: follows prompt intent, otherwise the first media item (video takes priority over images)
  4. Video editing / extension (2.5): the output ratio follows the input video being edited or extended
  5. The actual ratio used is returned in the task response’s ratio field
ratio only accepts the 7 enum values above — passing e.g. "2:1" returns an InvalidParameter error (verified), as does a duration outside the supported range (4-30 on 2.5, 4-15 on the 2.0 family). Neither is billed.2.5 additionally forces ratio: adaptive on three task types: first-frame / first+last-frame generation, video editing, and video extension. Passing a concrete aspect ratio there returns InvalidParameter.TaskTypeConstraint at submission time (measured: a synchronous 400, not a task that fails minutes later).

​
Best Practices

1

Pick the model by output needs

Start by asking whether you need what only 2.5 has: 30-second clips, 30 reference images, standalone audio references, mov output, or explicit video-edit/extend task types. If you need any of them, pick doubao-seedance-2-5-260628 (roughly 1.5× the standard model’s price, same group as the 2.0 family). If not, stay on the 2.0 family: for batch production and cost-sensitive workloads use the lite model doubao-seedance-2-0-mini-260615 (about half the standard price and the fastest generation, capped at 720p); for 1080p or top quality use the standard model; fast is the middle ground.
2

Ingest images and video to asset IDs first

When a request carries media, create-task latency is dominated by uploading that media: inline Base64 or a large image URL stretches submission from about a second to tens of seconds, or into a read timeout. Ingest the media first, get an asset:// asset ID, and reference that — the request body shrinks to a few dozen bytes, the task ID comes back immediately, and content checks move up to ingest time. See Asset-First Workflow.
3

Use adaptive to avoid cropping

For image-to-video keep the default adaptive so the model matches your source image’s ratio. Lock 9:16 (portrait) or 16:9 (landscape) only when the target platform demands it.
4

Duration is your cost dial

Cost scales strictly with length. Validate prompts with 5 s clips first, then scale up (2.5 goes to 30 s, which costs 6× a 5 s clip). On 2.5, always pass duration explicitly — its default is -1, so omitting it lets the model choose, and in testing it chose 10 seconds, doubling the cost.
5

Turn audio off when you don't need it

generate_audio defaults to true. Pass false for silent footage you plan to score yourself.
6

Quote dialogue for better voice-over

Put spoken lines inside double quotes in the prompt — the model generates matching voices automatically.
7

Add Accept-Encoding: identity in HTTP clients

The gateway labels responses content-encoding: gzip while the body is uncompressed; auto-decompressing clients such as Python requests raise ContentDecodingError. Adding the Accept-Encoding: identity header avoids this (curl is unaffected).
8

Poll every 15-30 s and download immediately

Tasks typically finish in 2-5 minutes. content.video_url is a signed link valid for 24 hours — copy the file to your own storage as soon as the task succeeds.
9

Chain clips with return_last_frame

Set return_last_frame: true to get a watermark-free last-frame png, then use it as the next task’s first frame to build continuous multi-clip videos.

​
Error Codes & Retries

Client recommendations:
  • 30-60 s request timeouts are enough for create/poll calls (the wait happens on the task side)
  • Poll every 15-30 s with an overall budget of 15+ minutes (longer for 1080p / 15 s tasks)
  • Apply exponential backoff on 5xx and timeouts (2 retries)
  • Log the task id and the x-request-id response header for troubleshooting

​
FAQ

It is submission that is slow, not generation. Your media has to travel upstream to APIYI, then be forwarded to Volcengine and decoded and validated there — the task ID only comes back after all of that. With inline Base64 or a large image URL, that step goes from about a second to tens of seconds, and even a 60-second client read timeout may not be enough. Generation itself normally takes 2–5 minutes, which is the provider’s normal speed and unrelated to submission.The real fix is to ingest the media first and reference it as an asset:// asset ID, which drops the request body from megabytes to a few dozen bytes. For the latency breakdown, migration steps, and how to tell whether a task was created after a timeout, see Asset-First Workflow.
Pick by the capability you need — 2.5 is not an automatic default. It runs at roughly 1.5× the 2.0 standard model (720p/5s: $1.3721 vs $0.9074), which mirrors the gap in Volcengine’s own list prices.Go to 2.5 when you need clips longer than 15 seconds (2.5 reaches 30), more than 9 reference images (2.5 takes 30), mov output, or video editing / extension with parameter errors surfacing at submission time (omni_reference_task_type). 2.5 also allows audio as the only reference, where 2.0 requires an image or video alongside it.Stay on the 2.0 family when your clips are under 15 seconds — the standard model also supports 1080p and sits in the same flagship quality tier; for cost-sensitive batch production use mini at about half the standard unit price and the fastest generation. The 2.0 family is not being retired.Endpoint, auth, and request shape are identical across generations, and so is the group — switching means changing one field: model.
1080p works (measured: 1920×1080 renders normally). 4k does not — sending "resolution": "4k" returns a synchronous 400 (not billed).One easily missed difference: 2.5 encodes 1080p as H.265 (hvc1), while 480p and 720p use H.264 (avc1). H.265 files are smaller, but older players, some browsers, and certain editing suites handle it less reliably than H.264 — confirm your downstream pipeline can decode it before distributing 1080p.
Both are “omni reference” tasks, triggered by a reference video in content plus the intent expressed in your prompt. Pass omni_reference_task_type explicitly so errors surface early:
  • Video editing: omni_reference_task_type: "edit", at least one role: "reference_video", ratio must be adaptive and duration must be -1, and the source video must run 4-30 seconds. The prompt needs an editing verb (add, remove, delete, change, replace). Output ratio and duration follow the input video — and the duration can be fractional (a measured run returned 16.709 seconds).
  • Video extension: omni_reference_task_type: "extend", again with a reference video and ratio set to adaptive. The prompt needs an extension verb (extend, continue).
Refer to assets positionally in the prompt — @video1, @image1 — in the order you passed them. Invalid parameters return a 400 at submission (InvalidParameter.TaskTypeConstraint) rather than failing the task minutes later.
Passing "output_format": "mov" returns a QuickTime container (H.264 + yuv444p chroma + PCM audio) with higher colour and luminance fidelity — suited to grading, keying, and compositing, and recommended officially as both input and output for video editing / extension work. The default is mp4, which has the broadest compatibility.Note that mov uses professional codecs some players cannot open (VLC, mpv, ffplay, and IINA on macOS all handle it). For direct web or mobile distribution, stay on the default mp4.
The most common Seedance error, and nine times out of ten the Token has the wrong group enabled. The error names the group you are currently on, for example:
Match your model to its group and enable it in Token Settings:One Key with SeeDance2 enabled reaches all four models. The billing model must also be Pay-as-you-go Priority or Pay-as-you-go — Pay-per-request Tokens cannot route.
The gateway’s content-encoding: gzip header does not match the actual body encoding. Symptoms include ContentDecodingError, a truncated non-JSON body (e.g. the leading {" is lost and you only get id":"cgt-xxx"}), or intermittent 400s. Add "Accept-Encoding": "identity" to your request headers; curl and browser fetch are unaffected.
generate_audio defaults to true (verified): the model adds voice, sound effects, and background music automatically. Pass "generate_audio": false explicitly for silent output.
On success the URL is at content.video_url in the poll response (not top-level). It is a signed link valid for ~24 hours — download and re-host it immediately. The task_id itself remains queryable for 7 days.
The state machine is queued → running → succeeded / failed / expired. The success state is succeeded, not completed — an easy mistake when migrating from other video APIs.
No. Neither Seedance 2.5 nor the 2.0 family accepts reference images/videos containing real human faces (upstream content safety). Alternatives: reuse face-containing output generated by Seedance models within the last 30 days, use the platform’s preset virtual avatars (asset:// IDs), or use licensed face assets.
No. Virtual-avatar ingest, real-person verification, and the rest of the private asset library are free with the Seedance 2.0 API on APIYI — no annual fee. Officially this capability is a separately purchased add-on: a six-figure CNY annual contract for customers without a framework agreement (Volcengine also sells it directly). We value long-term users, so that cost is already baked into our API price; for customers making normal use of the SD2 API, there is no extra charge at normal business volumes. See Asset Library for usage.
Parameter rejections (HTTP 400) are not billed (verified). Billing is pre-charged on submit and settled on completion, so your balance fluctuates briefly — reconcile against call logs.
tokens ≈ duration(s) × width × height × 24 / 1024, verified to within 0.1%. Every ratio in a tier has the same pixel area (720p 16:9 and 9:16 both cost 108,900 tokens per 5 s) — landscape, portrait, and square all cost the same.
Price and speed go mini < fast < standard (720p/5s nominal on-platform: about ¥3.16 / ¥5.08 / ¥6.35). Pick mini for batch production and cost-sensitive workloads — about half the standard price and the fastest generation (measured 2026-07: ~1.5-2.5 min for 5 s @720p). Pick standard for maximum detail, and fast as the middle ground. Both mini and fast cap at 720p — requesting 1080p returns a 400 parameter error (not billed).If 1080p is all you need, the 2.0 standard model already has it — no need to upgrade for that alone. Go to 2.5 for 30-second clips, more than 9 reference images, mov output, or video editing/extension (roughly 1.5× the standard model’s price, same group as the 2.0 family).
The model picks its own length (4-30 s on 2.5, 4-15 s on the 2.0 family) and bills by actual output. The final length is returned in the task’s duration field.Note that -1 is the default on 2.5 (the 2.0 family defaults to 5 seconds) — omitting duration therefore opts you into model-chosen length, and a measured 2.5 request with no duration returned a 10-second clip, exactly twice the cost of 5 seconds. Pass duration explicitly if cost predictability matters.
No. frames and camera_fixed are Seedance 1.x parameters — supported by neither Seedance 2.5 nor the 2.0 series. Use whole-second duration instead.
No — they are three mutually exclusive modes: first+last (2 images with required first_frame/last_frame roles), first frame (1 image), and multi-modal reference-to-video (image role reference_image). To approximate “first/last frame + reference”, use reference mode and designate a frame via the prompt.Reference limits differ by generation: 2.5 takes 30 images + 10 videos + 10 audio clips, and audio may stand alone; the 2.0 family takes 9 images + 3 videos + 3 audio clips and needs at least 1 image or 1 video alongside any audio.
The SeeDance2 group has ample concurrency with no queuing (15 simultaneous tasks all ran immediately in our test). Contact sales for larger sustained workloads.
Keep prompts under ~500 Chinese characters or ~1000 English words — longer prompts dilute detail. Supported languages: Chinese, English, Japanese, Spanish, Portuguese, Indonesian. Describe subject + action + camera movement + lighting/style.
Seedance is one of the few first-tier 2026 video model families that outputs synchronized audio by default. Combined with same-price aspect ratios and 2.5’s 30-second ceiling, it is a strong primary channel for short-video and e-commerce asset production. To compare alternatives, the same Token (with extra groups enabled) can call Sora 2, VEO 3.1, and Wan2.7 directly.