Last updated: August 23, 2026
Seedance 2.0 is ByteDance’s multimodal AI video model, launched in China on February 12, 2026 and now reachable through consumer apps and several APIs. This Seedance 2.0 review is built from the model card, ByteDance Seed’s launch materials, and the BytePlus ModelArk documentation rather than from hands-on testing. The aim is to show what the model is documented to do, where ByteDance says it falls short, and whom it suits.
ByteDance added 1080p and then 4K output to the API model in April and June 2026, released Fast and Mini tiers, worked through a copyright dispute with Hollywood studios, and then shipped Seedance 2.5 as a successor. Every specification, price, and date below comes from an official ByteDance or provider document or from cited reporting, and where those sources disagree, we say so.
- Seedance 2.0 accepts text, images, audio, and video as inputs (up to 9 images, 3 video clips, and 3 audio clips per request) and outputs 4 to 15 seconds of video with audio generated in the same pass.
- Three API tiers exist: Seedance 2.0 (up to 4K at 10-bit), plus Fast and Mini (both capped at 720p).
- The model card reports a number-one ranking on Arena’s text-to-video and image-to-video leaderboards but also lists deformation artifacts, audio noise, and multi-speaker lip-sync errors as open problems.
- As of publication, neither the API nor the Dreamina and CapCut apps accept uploads containing real human faces.
- A 5-second 720p clip lists at USD 0.76 on BytePlus ModelArk for the standard tier and USD 0.38 for Mini, before limited-time discounts.
Table of contents
What Seedance 2.0 is, in ByteDance’s own terms
ByteDance Seed’s launch post, dated February 12, 2026, describes Seedance 2.0 as a model that “supports four input modalities: text, image, audio, and video.” It accepts up to 9 images, 3 video clips, and 3 audio clips alongside natural-language instructions and produces 15-second multi-shot output with dual-channel audio. The model page frames this as a unified audio-video joint generation architecture, where references control performance, lighting, shadow, and camera movement.
The technical report, published as arXiv 2604.14148 on April 15, 2026, is the closest thing to a formal model card. It states that the model generates 4 to 15 seconds of audio-video content at native resolutions of 480p and 720p, and that it was “officially released in China in early February 2026” on Doubao, Jimeng, and Volcano Engine under the model ID doubao-seedance-2-0-260128, with a Fast variant for low-latency use.
One point of honest tension: the paper claims only 480p and 720p native output, while the BytePlus ModelArk model list now shows the same model at 1080p and 4K. ModelArk’s product-update log resolves this by date: 1080p arrived in April 2026 and 4K in June 2026, both after the paper was written.
Seedance 2.0 review: the core capabilities
Omni-reference inputs
ByteDance’s lead feature is what its docs call omni reference: mixing several kinds of reference assets in one request. The BytePlus tutorial allows 0 to 9 images, 0 to 3 videos, and 0 to 3 audio files per generation, with two exclusions. “Text + audio” and “audio-only” inputs are not supported on the 2.0 series, so audio must always be paired with at least one image or video.
Prompts bind assets to roles using tags. BytePlus’s prompt guide suggests using the subject in @Image 1 as the main character, @Image 2 as a scene style reference, and @Video 1 for camera movement. Dreamina’s app uses the same @ labels, while some API code samples use bracket forms such as [Image 1]; both appear in official material.
Dreamina’s Seedance 2.0 page caps reference resources at 12 in total (9 photos plus 3 videos plus 3 audio clips would be 15), and fal.ai’s reference-to-video endpoint carries the same 12 cap. Treat 12 as the practical ceiling on those surfaces; our image-to-video prompt guide shows how to allocate the slots.
Multi-shot narrative
The model card describes “native, professional multi-shot narrative capability” and the ability to “directly reference text-based storyboards.” BytePlus recommends writing multi-shot prompts as an explicit Shot 1 / Shot 2 / Shot 3 sequence and warns users to specify only one type of camera movement per shot.
Because output tops out at 15 seconds, multi-shot work on Seedance 2.0 means packing two or three cuts into a short clip rather than building a scene. Seedance 2.5 extends single-pass generation to 30 seconds; our rundown of what changed in Seedance 2.5 covers that trade-off.
Native audio
Audio is generated jointly with video rather than added afterward. The model card claims “binaural audio capability” with simultaneous multi-track output covering background audio, ambient sound effects, and character narration. Language coverage is narrower than the successor model: the BytePlus API reference lists English, Spanish, Indonesian, Portuguese, and Japanese for the 2.0 series, against eleven languages for Seedance 2.5. See our explainer on Seedance audio and dialogue generation for more.
Editing and extension
Two further capabilities stand out. Video editing is described in the model card as “targeted modifications to specified clips, characters, actions, or plot elements.” Video continuation is exposed on ModelArk as an extension task that can stitch up to 3 clips into one coherent video. All three tiers share the full capability set, so choosing Fast or Mini costs you resolution, not features.
The three tiers: Seedance 2.0, Fast, and Mini
The BytePlus Seedance 2.0 series tutorial positions the tiers as Seedance 2.0 for highest quality, Fast for a balance of cost and speed, and Mini for best cost performance. The standard and Fast models were logged as new ModelArk releases in April 2026. Mini carries the version suffix 260615.
| Tier | ModelArk model ID | Output resolutions | Duration | Positioning |
|---|---|---|---|---|
| Seedance 2.0 | dreamina-seedance-2-0-260128 | 480p, 720p, 1080p (8-bit); 4K (10-bit) | 4–15 s | Highest quality |
| Seedance 2.0 Fast | dreamina-seedance-2-0-fast-260128 | 480p, 720p (8-bit) | 4–15 s | Balance of cost and speed |
| Seedance 2.0 Mini | dreamina-seedance-2-0-mini-260615 | 480p, 720p (8-bit) | 4–15 s | Best cost performance |
The release log says Fast inherits “the core functions and advantages of the seedance 2.0 model, with faster generation speed,” but never quantifies the speedup. On the consumer side, Dreamina describes Mini as “A lightweight video model that blends speed, consistency and price” and caps it at 12 references split as 6 photos, 3 audio clips, and 3 video clips.
Output specifications and limits
The table below consolidates the documented API specs. Most values apply across all three tiers; the 4K details apply only to standard Seedance 2.0.
| Specification | Documented value |
|---|---|
| Frame rate | 24 fps |
| Duration | 4–15 seconds, or -1 to let the model choose |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Output format | .mp4; 4K uses H.265/HEVC at 10-bit |
| Reference images | Up to 9, each under 30 MB |
| Reference videos | Up to 3, 2–15 s each, 15 s total, under 200 MB, 24–60 fps |
| Reference audio | Up to 3, 2–15 s each, 15 s total, under 15 MB, .wav or .mp3 |
| Rate limits | 600 RPM / 10 concurrent (enterprise); 180 RPM / 3 (individual); 4K: 15 RPM / 1 concurrent |
4K output is encoded in H.265/HEVC, which BytePlus warns some players and browsers cannot play directly, so budget for a transcode. The 4K limit of one concurrent task applies to enterprise and individual accounts alike, which rules it out for high-throughput pipelines.
Offline or flex inference is not available for any 2.x model either. A fuller breakdown lives in our resolutions and durations guide.
Arena leaderboard results
The model card’s headline benchmark claim is that Dreamina Seedance 2.0 at 720p “ranks #1 on both the Text-to-Video and Image-to-Video leaderboards” of Arena, formerly LMArena, “with Elo scores of 1450 (±15) and 1449 (±11) respectively.” It reports a 79-point lead over second-place veo-3.1-audio-1080p on text-to-video (our Seedance versus Veo 3.1 comparison has the specs) and a 29-point lead over grok-imagine-video-720p on image-to-video.
Those are strong numbers with two caveats. They are a snapshot from the April 2026 paper, and leaderboards move quickly: by June 2026, VentureBeat was reporting Alibaba’s HappyHorse 1.1 at number two on the separate Artificial Analysis Video Arena with 1,444 in both categories. ByteDance’s model page also cites an internal benchmark, SeedVideoBench-2.0, which cannot be compared against anything else.
Documented weaknesses, in the model card’s words
ByteDance is direct about what still goes wrong. The model card states: “Areas for improvement remain: minor deformation artifacts, motion plausibility in edge cases, high-frequency visual noise, audio distortion and noise, and lip-sync errors in multi-speaker scenes.” A separate passage adds that “there is still room for optimization in multi-subject consistency, text restoration accuracy, and the performance of complex editing tasks.”
In practice that covers dialogue with more than one speaker, scenes with several distinct characters, legible on-screen text, and involved editing instructions. The Chinese launch announcement was similarly candid, calling Seedance 2.0 far from perfect. Our guide to common Seedance errors and artifacts collects documented mitigations.
The real-face restriction
This is the limitation most likely to surprise new users. BytePlus’s API reference warns that the Seedance 2.0 series and Seedance 2.5 “do not support direct uploads of reference images or videos containing real human faces.” CapCut’s rollout notes restrict “the ability to make videos from images or videos that contain real faces” in the initial rollout, and Dreamina’s support account has said the system automatically blocks such uploads for compliance reasons.
ModelArk documents three sanctioned workarounds: reusing face-containing outputs the same account generated within the past 30 days, a preset digital character library, and authorized real-person assets under BytePlus’s usage rules. None lets you drop in a photo of yourself or a client, so anyone planning likeness-based work should read our breakdown of Seedance’s content rules first.
The IP controversy and the safeguards that followed
Seedance 2.0’s launch drew a copyright backlash within a day. On February 12, 2026, Motion Picture Association chairman Charles Rivkin said the service “has engaged in unauthorized use of U.S. copyrighted works on a massive scale” and called on ByteDance to “immediately cease its infringing activity.” Disney sent a cease-and-desist the next day calling the model a “virtual smash-and-grab of Disney’s IP,” and the MPA followed with a formal cease-and-desist letter on February 20. ByteDance said on February 16 that it was “taking steps to strengthen current safeguards,” and the planned mid-March global rollout, including BytePlus API access and a consumer app outside China, was paused.
The revised international release came with a package of controls: blocking real-face inputs, filters against copyrighted characters, visible watermarks and AI labels, C2PA Content Credentials, invisible watermarking, and testing by a third-party red-teaming partner. CapCut’s log shows the expansion reaching Europe on March 28, Japan on March 31, and the United States on April 7, 2026. On August 17, 2026, the MPA and ByteDance announced a Memorandum of Understanding establishing guardrails across Seedance, Seedream, TikTok, CapCut, and Dreamina, though The Next Web noted that neither side has spelled out what those guardrails actually are.
Watermark behavior differs by surface. On the ModelArk API the visible watermark parameter defaults to false, while C2PA Content Credentials are embedded by default for Seedance 2.0, Mini, and Fast outputs. On Dreamina, watermark-free exports and lip-sync support are listed as paid-plan features, and BytePlus’s terms prohibit removing watermarks or credentials from output. See our guide to Seedance watermarks and commercial rights.
Pricing snapshot
BytePlus ModelArk bills by tokens, where estimated consumption equals input video duration plus output duration, multiplied by output width, height, and frame rate, divided by 1,024, and only successful generations are charged. Standard Seedance 2.0 lists at USD 7.0 per million tokens without video input and 4.3 with it at 480p and 720p; Fast lists at 5.6 and 3.3, and Mini at 3.5 and 2.1. The worked examples below, for a 5-second 16:9 clip, are easier to compare.
| Tier | 480p | 720p | 1080p | 4K |
|---|---|---|---|---|
| Seedance 2.0 | USD 0.35 | USD 0.76 | USD 1.87 | USD 3.89 |
| Seedance 2.0 Fast | USD 0.28 | USD 0.60 | Not supported | Not supported |
| Seedance 2.0 Mini | USD 0.18 | USD 0.38 | Not supported | Not supported |
Two promotions are running. From August 7 through September 7, 2026, Mini is billed at 40% of list and Fast at 75% of list for 480p and 720p, roughly USD 0.03 and USD 0.09 per second at 720p.
Third-party hosts price differently (as of August 23, 2026). fal.ai quotes about USD 0.30 per second at 720p on its standard text-to-video endpoint, 0.2419 for Fast, and 0.682 at 1080p. Replicate bills USD 0.18 per output second at 720p for standard, 0.15 for Fast, and 0.09 for Mini.
On the consumer side, Dreamina’s promotional monthly plans list at USD 5, 11, and 42, with 225 free daily credits on the free tier. For provider-by-provider tables, see our Seedance API pricing comparison, or run your own numbers in the cost calculator.
Verdict by user type
Marketers and e-commerce teams
Seedance 2.0 suits product-led video where the hero asset is an object, a garment, or an environment. Omni reference locks a product’s look across shots with up to 9 images, the six aspect ratios include 9:16 for vertical placements, and Mini brings a 5-second 720p clip to USD 0.38 at list. The blockers are people: you cannot upload a founder or a model, and multi-speaker lip-sync is a documented weak spot. Check your surface’s terms too: Dreamina’s describe the service as generally for private, non-commercial use.
Filmmakers and narrative creators
Storyboard prompting, targeted editing, three-clip extension, and 4K 10-bit output add up to a notably filmmaker-oriented feature set. Against that, 15 seconds per generation is short, 4K runs one task at a time, and the model card’s artifact and motion-plausibility caveats bite hardest in long, complex shots. Serious sequence work probably belongs on Seedance 2.5, while 2.0 remains the only family member with a documented 4K API option.
Developers
The API story is mature: three model IDs with one capability set, a documented token formula, published rate limits, C2PA embedded by default, and hosting on BytePlus ModelArk, fal.ai, Replicate, OpenRouter, and Runway’s developer platform. Two constraints need planning: BytePlus’s terms state its model services are not available in the United States, and no offline tier exists for any 2.x model. Our Seedance API access guide maps the routes.
Hobbyists
Dreamina says Seedance 2.0 is live globally with free daily credits, which its documentation puts at 225, and watermarks are possible on free-tier exports. Chinese users have Doubao and Jimeng, where launch-week reporting described free daily video quotas for test users. Expect queues and a real-face block on anything you upload. Our guide to using Seedance for free lists every legitimate no-cost route.
Honest limitations and what we could not verify
This is a documentary review: we did not test Seedance 2.0 ourselves, and the quality judgments above are ByteDance’s own as printed in the model card. Any claim that the model is “best” at something should be read as a leaderboard claim with a date attached.
Several things could not be confirmed from official sources. There is no standalone ByteDance announcement for Seedance 2.0 Mini. The speed advantage of Fast and Mini is described only qualitatively in official documentation. Real-world generation times, queue lengths, and failure rates are not documented anywhere official.
Pricing is a moving target. The BytePlus figures are list prices already subject to two promotions, and Dreamina’s plan names and prices are flagged as promotional and regional. Finally, the guardrails in the August 2026 MPA agreement have not been described publicly by either party, so we cannot say how they change outputs compared with the March safeguards.
Frequently asked questions
Is Seedance 2.0 still worth using now that Seedance 2.5 exists?
Yes for some workloads. Seedance 2.5 extends single-pass generation from 15 to 30 seconds and raises the reference limit to 50 assets, but on the BytePlus API only Seedance 2.0 offers 4K output, and the Fast and Mini tiers list well below 2.5’s per-token price. Pick 2.0 for cost or 4K, and 2.5 for length and reference count.
Can Seedance 2.0 generate 4K video?
The standard dreamina-seedance-2-0-260128 model on BytePlus ModelArk supports 4K output, added in June 2026, encoded in H.265/HEVC at 10-bit color depth. Fast and Mini are limited to 480p and 720p. 4K requests are capped at 15 per minute and one concurrent task, and BytePlus warns some browsers cannot play HEVC files directly.
Does Seedance 2.0 allow uploading photos of real people?
No. BytePlus states that Seedance 2.0 series models do not support direct uploads of reference images or videos containing real human faces, and CapCut and Dreamina apply the same block. The documented alternatives are face-containing outputs the same account generated within the past 30 days, preset digital characters, or authorized real-person assets.
How much does Seedance 2.0 cost per video?
On BytePlus ModelArk a 5-second 16:9 clip without video input lists at USD 0.35 at 480p, 0.76 at 720p, 1.87 at 1080p, and 3.89 at 4K for the standard tier; Fast is 0.28 and 0.60 and Mini 0.18 and 0.38 at 480p and 720p. fal.ai charges about USD 0.30 per second at 720p and Replicate USD 0.18.
Seedance 2.0 is a documented leader on one major leaderboard, the only Seedance model with a 4K API option, and unusually flexible about mixing references; it is also capped at 15 seconds, blocked from real faces, and candid about its artifacts. For wider context, start with our complete Seedance 2.0 guide, then work through every working way to access Seedance and the multi-shot storyboard prompt templates.
Sources: Seedance 2.0 technical report (arXiv 2604.14148); BytePlus ModelArk Seedance 2.0 documentation (docs.byteplus.com); Motion Picture Association press releases (motionpictures.org); ByteDance Seed blog; CapCut and Dreamina pages.
