Seedance 2.5, launched by ByteDance in 2026, accepts up to 50 combined reference inputs per request, more than any earlier Seedance release. It suits creators inside Jimeng AI or Doubao Pro today, not developers wanting a scored, publicly documented video API.
Seedance 2.5 is ByteDance's video generation model, launched in 2026, that produces a single continuous 30-second clip in one generation pass with synchronized audio-video output. It supports multiple combined image, video, and audio reference inputs per request, aimed at short-form and marketing video creators rather than developers.
Provider: ByteDance · Family: Seedance 2
Input modalities: text, image, video, audio · Output: video, audio
About Seedance 2.5
Seedance 2.5 is ByteDance's AI video generation model, unveiled at the Volcano Engine FORCE conference on June 23, 2026, and launched publicly on July 31, 2026 inside Jimeng AI and Doubao Pro. It succeeds Seedance 2.0, which shipped roughly five months earlier, and sits inside ByteDance's Seed research organization alongside the original Seedance 1.0 model documented in a June 2025 arXiv paper. ByteDance describes the underlying architecture as a diffusion transformer with sparse attention, though no technical report for 2.5 had been published as of this writing, so treat that description as ByteDance's own characterization rather than an independently verified specification. No parameter count has been disclosed. The standout capability is duration: a single generation produces one continuous clip lasting thirty seconds, twice the length of the prior release, with an option to extend the output twice more for longer storytelling. ByteDance's own product page markets native 4K output at 10-bit color, but that figure has not been confirmed on any public pricing or specs page, and one independent tracker reports the top resolution may sit behind a specific access tier rather than shipping by default. Frame rate and supported aspect ratios were not published anywhere in official documentation found during research. Creators can attach up to fifty combined reference inputs per request, spanning images, video clips, audio, and white-model 3D blockouts, more than triple the prior release's cap. This lets a creator anchor camera movement, character motion, and performance blocking to a reference before generating the final shot. The model also produces audio and video together in the same pass rather than needing a separate dubbing step, and supports editing a specific region or moment of an already-generated clip without re-rendering the rest of it. As of August 9, 2026, direct API access through BytePlus ModelArk is still listed as coming soon rather than generally available. The model is reachable today only through ByteDance's own consumer apps, plus a handful of early third-party resellers including Kie.ai, Hedra, OpenArt, and Apiframe that wired up their own access layers ahead of the official API launch. No confirmed list of supported SDK languages, hosting regions, or self-hosting options exists for this release specifically. Secondary sources cite unconfirmed per-token pricing in Chinese yuan for API access, priced differently depending on whether the request includes video input, alongside a small worked cost example for a short low-resolution clip. These figures could not be confirmed against BytePlus's own live pricing console, which returned only navigation content with no extractable rates at the time of research, so exact numbers are omitted from this record's structured pricing fields rather than presented as confirmed facts. No independent third-party benchmark exists for this release as of August 2026. Artificial Analysis's Video Arena, the closest thing to a crowd-sourced standard for video model quality, lists the original Seedance 1.0 and Seedance 1.5 Pro but had not published a score for Seedance 2.5 at the time of research; one search result explicitly warned that any 2.5 Elo score circulating online should be treated as fabricated. Seedance 2.0 scored an Elo of 1,269 on that same arena, reportedly ahead of Veo 3, Sora 2, and Runway Gen-4.5, but that number describes the earlier model, not this one. On safety, ByteDance documented for Seedance 2.0 a block on generating video from real human face images following a deepfake controversy earlier in 2026, plus C2PA and invisible watermarking positioned around the EU AI Act's transparency deadline. Neither policy has been separately re-confirmed for this release in ByteDance's own materials, so buyers evaluating it for sensitive use cases should verify current policy directly with ByteDance or Volcano Engine rather than assume the earlier policy carries over unchanged. This model fits creators and marketing teams who want a single-pass, long-form clip without manual stitching, particularly inside ByteDance's own consumer ecosystem where it already runs live. It is a weaker fit today for engineering teams that need a benchmarked, generally available API to build a production pipeline against, since neither condition is met yet: no independent quality score exists, and first-party API access remains pending.
Pricing
BytePlus's own live pricing console returned no extractable rates at review time, so this record leaves per-token pricing unset instead of publishing an unconfirmed figure. Secondary trackers (CometAPI, Kie.ai) separately report roughly 70 CNY per 1M tokens without video input and 42 CNY per 1M tokens with video input for the eventual API tier.
Key Features
- Single-pass 30-second generation: One generation call outputs a continuous 30-second clip, the longest native single pass ByteDance has shipped under the Seedance name, with an optional extend step for longer sequences.
- 50-input reference budget: A single request can combine up to 50 image, video, audio, and 3D blockout references, giving tighter creative and camera control than earlier text-only prompting.
- Native audio-video joint generation: Produces synchronized audio alongside the video in the same generation pass, skipping the separate dubbing step most competing video models still require.
- Region and timestamp-level post-editing: Lets a creator revise one region or moment of an already-generated clip locally, without re-rendering the rest of the output.
- Reference-video motion capture: Transfers motion and performance from an attached reference clip onto the generated scene, per ByteDance's own model page.
Pros
- Produces a single continuous clip per generation instead of stitching multiple shorter clips together, the difference ByteDance highlights most in its own launch materials.
- Combines image, video, audio, and 3D blockout references in one request, giving tighter creative and camera control than text-only prompting alone.
- Generates audio and video together in a single pass instead of requiring a separate audio-generation step afterward.
Cons
- No independent benchmark exists yet: Artificial Analysis's Video Arena had not published a score for this release as of August 2026, so quality claims rest on ByteDance's own marketing.
- Direct BytePlus ModelArk API access remained listed as coming soon as of August 2026, so first-party programmatic integration was not yet available.
- Exact resolution tiers are ambiguous: ByteDance markets native 4K, 10-bit output, but independent trackers say the sharpest setting could require a paid upgrade rather than shipping by default.
Frequently Asked Questions
How much does Seedance 2.5 cost?
Confirmed pricing was not available when this record was written: BytePlus's pricing console showed no extractable rates, and direct API access remained listed as coming soon. Third-party trackers estimate the eventual API rate near 70 CNY without video input and 42 CNY with video input per million tokens, but treat those figures as unverified until BytePlus publishes its own numbers.
How does Seedance 2.5 compare on benchmarks to Veo 3.1, Kling, or Sora 2?
No independent benchmark score exists for this release as of August 2026, and any 2.5 Elo figure circulating online has been flagged as fabricated. Its predecessor, Seedance 2.0, sits at Elo 1,269 on that same arena, above three rival video models named in Artificial Analysis's rankings, but that figure belongs to the older model and cannot be assumed to carry over.
Is Seedance 2.5 open source or proprietary?
Seedance 2.5 is proprietary, with no open weights or license. Access currently runs through ByteDance's consumer apps and a few early resellers, such as Kie.ai, Hedra, OpenArt, and Apiframe, since BytePlus has not yet flipped its own API to generally available.
Does Seedance 2.5 train on user data?
ByteDance has not published a data retention or training-on-inputs policy specific to this release. Its predecessor added C2PA and invisible watermarking after a deepfake controversy tied to real-face uploads, but ByteDance has not restated either safeguard for this newer model.
Who is Seedance 2.5 best for and who should avoid it?
It suits social and marketing video makers working inside Jimeng AI or Doubao Pro, where the model already ships. Developers who need a scored, generally available API should look at an already-evaluated option like Veo 3.1 or Kling instead, at least until Seedance 2.5 gets an independent score and its own API leaves preview.