Wan 2.7 vs Seedance 2.0: Which Video Model Fits Your Brief?

Both Wan 2.7 and Seedance 2.0 are current video-generation options. Start with Wan 2.7 when you need a 720p or 1080p clip and its text-to-video, image-to-video, reference-to-video, or editing operation matches your inputs. Start with Seedance 2.0 standard when native 4K is required, or with the Seedance 2.0 series when you need its documented combination of image, video, and audio references. At 720p, Seedance Fast and Mini also offer lower estimated generation costs than the standard tier.

That is a choice based on documented requirements, not a claim that one model consistently produces better motion, faces, or finished ads. This comparison uses the current Alibaba Cloud Model Studio and BytePlus ModelArk documentation checked on October 12, 2026. We did not generate comparison clips or measure latency, acceptance rates, or actual bills.

Choose the operation before choosing the brand

A reference image can mean an opening frame, a subject to preserve, or an example of visual style. Those roles are different. Decide which role your asset must play before submitting it.

These choices follow the Wan T2V, I2V, R2V, and video-editing contracts and the Seedance 2.0 series tutorial. They describe API capabilities, not guaranteed creative results or every consumer app that uses these brands.

Wan 2.7 also has image products, which remain useful for preparing compositions and reference packs. Their existence no longer means that Wan 2.7 is image-only. A still-image API and a video API are separate choices, even when the version number matches.

What each video workflow actually accepts

Wan: match the media to T2V, I2V, R2V, or editing

For text-to-video, wan2.7-t2v accepts a prompt and an optional audio file, with 720P or 1080P output, five aspect ratios, and an integer duration of 2–15 seconds. Use the current resolution and ratio fields. The older size field and shot_type setting are not the Wan 2.7 T2V contract; describe the shot structure in your prompt instead. Alibaba's T2V reference supplies the current fields and regional endpoints.

For image-to-video, the documented Wan 2.7 combinations are a first frame, a first frame plus last frame, or a first video clip to continue. The image-based combinations can also take driving audio. Each media type can appear only once. This makes it a suitable starting point when your approved opening and ending compositions matter more than assembling a large reference library.

The Wan I2V contract allows 2–15-second output at 720P/1080P. Output proportions follow the first frame or first clip, with possible rounding drift; use a correctly shaped input and allow for a final crop if exact dimensions matter. For continuation, an input clip must be 2–10 seconds, and duration means the total output including that clip. For example, an 8-second first clip with duration: 15 leaves 7 seconds for new footage, not 15.

For reference-to-video, wan2.7-r2v accepts at least one reference image or reference video, with a combined maximum of five, plus an optional first frame. A subject reference must contain a single character. Images and videos are numbered separately in the prompt: Image 1 and Video 1 refer to their order within their own media type. If reference input includes video, output duration is 2–10 seconds; without video, it is 2–15 seconds. A first frame overrides the requested aspect ratio. These are the R2V operation's rules, not a universal nine-image limit for every Wan video request.

For video editing, wan2.7-videoedit takes one 2–10-second source video and up to four reference images. Use it for an instruction such as changing a setting or replacing an element in existing footage. Omit duration, or set it to 0, to keep the source duration; a positive value truncates from the start. It does not extend the video. audio_setting: "origin" retains the original sound, whereas the default auto may generate new sound according to the instruction. The editing reference documents those distinctions.

Seedance: choose the tier, then the reference mode

BytePlus lists three current Seedance 2.0 model IDs:

• Standard: dreamina-seedance-2-0-260128
• Fast: dreamina-seedance-2-0-fast-260128
• Mini: dreamina-seedance-2-0-mini-260615

All three support text-to-video, first-frame and first-and-last-frame generation, omni reference, editing, extension, and video with audio. All output 4–15-second clips. Standard supports 480p, 720p, 1080p, and 4K; Fast and Mini support 480p and 720p. The six listed ratios include 21:9 as well as the usual landscape, portrait, and square formats. Standard's 4K output is 10-bit H.265/HEVC, so delivery compatibility is part of the choice. These specifications come from the current BytePlus model matrix.

Omni reference lets you assign different jobs to media: an image can guide appearance, a video can guide movement, and audio can guide sound. The tutorial lists up to nine images, three videos, and three audio files, with combined media types supported. You do not need to fill every slot. Text-plus-audio alone and audio-only input are excluded. If the opening and ending images must serve as actual endpoints, use the first-and-last-frame feature with first_frame and last_frame roles rather than only mentioning them in an omni-reference prompt. The tutorial's reference guidance makes that distinction explicit.

Seedance editing can add, remove, or replace elements using reference footage and instructions. Extension can continue forward or backward, or connect up to three clips. A single-clip extension usually returns the new tail footage unless the prompt requests inclusion of the original; connecting multiple clips includes original and generated content. Confirm which output you need before comparing it with Wan continuation, which counts the source clip inside its total output duration.

Audio and faces can decide which workflow is usable

Use driving audio when the recording should drive the performance; use a voice reference when you want a timbre. In Wan I2V, driving audio guides lip-sync and action timing. In Wan R2V, reference_voice provides the character's voice characteristics and does not determine the spoken content. It overrides audio in a reference video when both are supplied. Treating these as interchangeable can produce the wrong dialogue even when the request is valid. See the I2V audio rules and R2V voice rules.

Wan T2V and image-based I2V also handle the length of supplied audio explicitly: excess audio is trimmed, while a recording shorter than the video leaves the remaining portion silent. Match the recording to the intended clip instead of assuming the model will finish it. Seedance omni reference supports voice timbre and audio-content guidance, and its models can generate audio with the video. Neither documented support nor a demo proves that every requested line or sync point will be correct; listen to the result as part of acceptance.

For portrait workflows, BytePlus ModelArk does not accept direct reference-image or video uploads containing real human faces for Seedance 2.0. Its supported routes are qualifying trusted model outputs, preset digital characters, or verified and authorized real-person assets. Trusted face-containing outputs must be original, generated on ModelArk under the same account, from the eligible models, and within 30 days. Cross-platform files, edits, compression, or forwarding can invalidate that trust. Output moderation still applies. The portrait-asset guide documents the eligibility and asset-ID routes.

This changes the useful “use both” advice. Preparing a face-containing image in Wan does not automatically make it an eligible Seedance reference. A non-face product, environment, or storyboard pack can still be a sensible handoff, provided it meets the destination's input requirements. For a specific person, establish the permitted portrait-asset route before investing in the reference pack.

Compare a generation budget with the same output target

For ordinary five-second, 16:9 generation without input video, the documented USD amounts give a useful starting budget:

Alibaba's Wan T2V model card lists these original rates under Singapore / International. BytePlus's pricing page supplies the five-second examples. Amounts exclude promotions, free quotas, taxes, and post-production. They do not establish equal picture quality, frame rate, codec, or acceptance rates. Different regions, resellers, and operations need their own prices.

As a planning example, twelve such 720p generations would be USD 6.00 on the named Wan T2V route, approximately USD 9.12 on Seedance standard, USD 7.20 on Fast, or USD 4.56 on Mini. This is arithmetic from the published rates and estimates, not a bill we incurred. If your final requirement is 1080p, choosing Mini because its 720p estimate is lower changes the deliverable; include any upscaling work separately rather than calling the two prices equivalent.

BytePlus bills video generation by tokens. Its estimate uses (input video duration + output video duration) × output width × output height × output frame rate / 1024; actual consumption is returned as usage.completion_tokens. Input-video jobs also have a minimum token charge that depends on resolution, ratio, and output duration. Consequently, a lower with-video token rate does not make a reference-video job cheaper. The publisher's standard-tier example for five-second, 16:9, 720p output with 2–15 seconds of input video is approximately USD 0.84–1.86, versus USD 0.76 without input video. See the pricing conditions and example tables.

Wan also distinguishes operations in billing. T2V counts output duration; R2V and video editing count input plus output video duration. I2V continuation counts the total returned video, including its first clip. Do not apply the T2V table above to a reference-video or editing job merely because it ends in a five-second video. The relevant R2V, editing, and I2V documentation defines what is counted.

Start from the correct official API path

Wan 2.7 through Alibaba Cloud Model Studio

1. In Model Gallery, confirm the exact model and its region. Your model, endpoint, and API key must share that region.
2. Open the reference for the required operation. For Wan 2.7 T2V, use the new HTTP protocol; the legacy Python/Java T2V SDK section explicitly supports Wan 2.6 and earlier, not 2.7. This restriction does not describe every Wan 2.7 operation's SDK support.
3. Set the output requirements explicitly, create one asynchronous task, and retain its task ID. Poll that task for completion instead of submitting another generation to check progress.
4. On SUCCEEDED, save the video bytes, the requested settings, and the returned usage. The task ID and temporary result URL expire after 24 hours.

For Singapore, the T2V reference gives the workspace endpoint below. Replace {WorkspaceId} with your workspace ID, authenticate with the matching regional key, and send Content-Type: application/json and X-DashScope-Async: enable.

POST https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/api/v1/services/aigc/video-generation/video-synthesis
GET  https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/api/v1/tasks/{task_id}

A T2V request body for a small initial generation can look like this. It is an illustrative request, not an executed test:

{
  "model": "wan2.7-t2v",
  "input": {
    "prompt": "A paper sailboat crosses a shallow puddle. One continuous close shot follows it from left to right. Soft rain sounds, no dialogue."
  },
  "parameters": {
    "resolution": "720P",
    "ratio": "16:9",
    "duration": 5,
    "prompt_extend": false
  }
}

That body generates from text. To animate approved endpoint images, use the I2V media contract instead; to borrow a subject or footage, use R2V. Changing only the model name does not make incompatible inputs valid.

Seedance 2.0 through BytePlus ModelArk

Start at the Seedance 2.0 tutorial, activate the required model, and use its matching example. The activation conditions currently include one of: an account balance greater than USD 30, a qualifying savings plan at the USD 30 tier or above, or a purchased resource pack with available quota. These are provider prerequisites, not the cost of one clip. The current matrix includes individual and enterprise users, so “enterprise-only beta” is no longer an accurate blanket description; it also does not promise availability to every account or location.

The tutorial uses https://ark.ap-southeast.bytepluses.com/api/v3 as the ModelArk base URL. The create-task endpoint is POST /contents/generations/tasks; the model-specific tutorial demonstrates client.content_generation.tasks.create(...) and polling with client.content_generation.tasks.get(task_id=...). Save the returned task ID, wait for succeeded, and handle failed before attempting a revised generation.

Choose the exact 2.0 ID listed above, set resolution, ratio, duration, and whether to generate audio, and use the tutorial's example for your chosen mode. For omni reference, image/video/audio assets have their documented reference roles; for endpoint images, use first/last-frame roles. The generic create-task page currently showcases Seedance 2.5, so its sample is not a drop-in 2.0 recipe. Do not combine extra mode-specific fields just because their labels appear on the same API page.

Save the generated file promptly: BytePlus's portrait workflow guide also states that the original generated video URL lasts 24 hours. Keeping a durable copy and qualifying for the 30-day trusted portrait-input window are separate requirements.

Keep the reference pack, measure the accepted result

Preparing the look before generating motion remains useful. For a non-face scene, approve the framing, subject details, opening and ending compositions, and sound intent first. A prepared pack can feed either model's suitable operation; it does not have to be a Wan-to-Seedance pipeline. Use fewer, clearly assigned references rather than adding files with conflicting roles.

When choosing between two otherwise compatible routes, give both the same brief and output target, then judge the files against your actual delivery requirements. Define those requirements before viewing results: the requested action occurs, the subject stays recognizable, critical product details survive, dialogue is intelligible, the endpoints work in your edit, and the file meets the required dimensions and playback format. These are suggested acceptance criteria, not test results from this article.

Record every charged generation, including successful outputs you reject. Cost per accepted clip = total recorded generation charges ÷ accepted clips. If none are acceptable, there is no finite accepted-clip average. Add editing labor or other services only if you track them consistently for both routes. BytePlus says moderation failures are uncharged, but an API-successful video that fails your creative brief is a different case. Its pricing policy distinguishes successful generation from failure.

This is where your own comparison can change the first choice. A cheaper request is useful when it produces an acceptable deliverable with less total work. Once you have actual results, use those results to decide whether to continue, adjust the brief, or switch operations. For help constructing that brief after selecting Seedance, the Seedance prompt guide explains reference roles, shot descriptions, and targeted repairs by model version.

Frequently asked questions

Does Wan 2.7 generate video, or is it only an image model?

It generates video through currently documented T2V, I2V, R2V, and editing APIs. Wan 2.7 image products also exist, but they do not define the whole version family. Start with the video operation that matches your inputs.

Which is cheaper: Wan 2.7 or Seedance 2.0?

For five-second 720p generation without input video, Alibaba's Singapore/international Wan T2V rate gives USD 0.50; BytePlus estimates USD 0.76 for Seedance standard, USD 0.60 for Fast, and USD 0.38 for Mini. These provider-specific amounts compare generation budgets, not accepted-clip costs. Different input operations and actual usage can change the result. See the Wan price card and BytePlus estimates.

Can both models use first and last frames and generate sound?

Yes. Wan I2V documents first-and-last-frame generation with optional driving audio; the Seedance 2.0 matrix lists first-and-last-frame generation and generated audio across standard, Fast, and Mini. Use the explicit endpoint-image mode rather than assuming ordinary reference images will pin the endpoints. See Wan I2V and Seedance 2.0.

Can I pass a face generated in Wan directly into Seedance 2.0?

Do not assume it qualifies. BytePlus's trusted portrait-output route is limited to eligible original outputs from ModelArk under the same account within 30 days. Cross-platform outputs are excluded from that trust route. Use an applicable preset, trusted output, or verified authorized-person asset instead, as described in the portrait guide.

Which one should I choose for a finished video?

Choose by the constraint you cannot relax: standard-tier Seedance for native 4K, Wan for a documented 2–3-second generation, or the operation that fits your reference/edit/continuation job. For a compatible 720p or 1080p brief, use the budget comparison to choose your first attempt and judge the actual file against your delivery criteria. Documented features alone do not establish a universal quality winner.