Seedance 2.0

Overview

Seedance 2.0 is ByteDance Seed's video generation model, launched on February 12, 2026 in China. ByteDance's own one-line description on its model directory is the tightest summary available:

"Adopts a unified multimodal audio-video joint generation architecture that supports text, image, audio, and video inputs, leading to the most comprehensive multimodal content reference and editing capabilities in the industry."

The launch post opens by calling it "our next-generation video creation model" and states that "built with a unified multimodal audio-video joint generation architecture, Seedance 2.0 supports four input modalities: text, image, audio, and video." Against its predecessor, ByteDance claims only that "compared with Version 1.5, Seedance 2.0 delivers a substantial leap in generation quality."

The distinguishing architectural claim is joint generation. Seedance 2.0 does not render silent footage and dub it afterwards — video and audio are produced together by one model. That, plus the four-modality input surface, is what ByteDance is selling. The technical report, posted to arXiv on April 15, 2026 with a team of 200+ contributors, describes "a unified, highly efficient, and large-scale architecture for multi-modal audio-video joint generation" which "integrat[es] one of the most comprehensive suites of multi-modal content reference and editing capabilities available in the industry to date."

The model became notorious almost immediately. Viral clips depicting recognisable actors and studio-owned characters drew cease-and-desist letters from five major Hollywood studios and then from the Motion Picture Association, and ByteDance paused the planned mid-March global rollout. That story is unresolved, and it is covered in Limitations rather than buried.

Seedance is the video half of ByteDance's generative media stack; Seedream 5.0 is the image half.

A note on naming

Seedance 2.5 has not shipped. It was announced on June 23, 2026 at the Volcano Engine FORCE conference in Beijing (Volcano Engine's own event page confirms the 2026.06.23–06.24 dates and the venue, though it names no model versions). A wave of SEO posts since then describe Seedance 2.5 as released, shipped, or "out now."

Four primary checks say otherwise, as of July 8, 2026:

1. ByteDance's own Dreamina landing page for the model reads: "Seedance 2.5 is coming soon!"
2. ByteDance Seed's model directory lists Seedance 2.0, Seedance 1.5 pro, and Seedance 1.0 — no 2.5.
3. There is no Seedance 2.5 post on the ByteDance Seed tech blog.
4. No doubao-seedance-2.5 model ID appears in Volcano Engine's model or pricing documentation.

Capability figures circulating for 2.5 — 30-second clips, 50 reference inputs, 4K output, local re-draw editing — trace to conference coverage and to ByteDance's own "coming soon" marketing page. They describe an unreleased product. Do not plan against them.

Capabilities

• Joint audio-video generation — one model produces synchronised video and audio. The launch post says Seedance 2.0 "has also enhanced its audio capabilities by integrating two-channel stereo technology, enabling high-fidelity, immersive sound generation."
• Four input modalities — text, image, audio and video, in combination. Up to 3 video clips, 9 images, and 3 audio clips as references in a single generation, per the technical report.
• Video editing — the launch post describes "targeted modifications to specified clips, characters, actions, and storylines."
• Video extension — generating continuous follow-on shots from a user prompt.
• Multi-shot narrative output — up to 15 seconds of multi-shot footage in one pass.
• Physical accuracy and controllability — ByteDance's headline improvement areas over 1.5, alongside visual realism. It singles out multi-participant competitive sports scenes as a case earlier models handled badly.
• Director-level control — the product page claims control over performance, lighting, shadow and camera movement.

Technical Specifications

• Developer: ByteDance Seed
• Architecture: unified multimodal audio-video joint generation. The launch post credits "the efficiency of a sparse architecture" alongside world knowledge and multimodal joint training; DeepLearning.AI's The Batch likewise notes that ByteDance "characterizes the architecture as 'sparse.'" ByteDance publishes no further architectural detail.
• Parameter count: not published by ByteDance. This page does not state one.
• Input modalities: text, image, audio, video
• Output: video with synchronised audio
• Native output resolutions: 480p and 720p (technical report)
• Duration: 4–15 seconds (technical report)
• Aspect ratios: 6 (CapCut's newsroom states "6 aspect ratios" without enumerating them; ByteDance's technical report does not list them)
• Reference input limits: up to 3 video clips, 9 images, 3 audio clips
• Variants: Seedance 2.0, and Seedance 2.0 Fast (accelerated, low-latency). ByteDance's Dreamina surface additionally markets a Seedance 2.0 Mini tier ("lower cost, fast batch video generation") that the technical report does not describe.
• Volcano Engine model ID: doubao-seedance-2-0-260128 (the 260128 suffix follows ByteDance's YYMMDD date-stamp convention); doubao-seedance-2-0-fast-260128 for the Fast variant
• arena.ai model string: dreamina-seedance-2.0-720p
• Weights: closed. No Hugging Face checkpoint, no self-hosting.

A resolution caveat worth stating plainly: the arXiv report says "native output resolutions of 480p and 720p." Dreamina's consumer page advertises "720p–1080p", and BytePlus publishes a 1080p pricing tier. ByteDance's paper does not describe 1080p as a native resolution, and the arena.ai entry that tops the image-to-video board is explicitly the 720p build.

Use Cases

• Short-form social video — the CapCut and Dreamina surfaces target this directly, at up to 15 seconds and six aspect ratios including 9:16.
• Short drama and micro-series production — the use case Chinese coverage of the pricing announcement fixated on, given a sub-¥15 cost for a 15-second shot.
• Reference-driven ad creative — supply product stills, a reference clip and a voice track, and generate variants against them.
• Targeted video editing — modifying a specified character, action or storyline inside existing footage rather than regenerating the clip.
• Shot extension and previsualisation — extending a shot with continuous motion, useful for storyboarding before committing budget.
• Sound-synchronised sequences — scenes where the audio must land with the action, generated in one pass rather than composited.

Performance / Benchmarks

ByteDance publishes no absolute numbers. Its product page shows radar charts asserting that Seedance 2.0 "leads across multiple dimensions" on its internal SeedVideoBench-2.0 evaluation, covering text-to-video, image-to-video and multimodal tasks, without printing the scores. The technical report's own summary is deliberately modest: the model "has demonstrated performance on par with the leading levels in the field" in "both expert evaluations and public user tests."

The useful numbers are third-party. These come from two different organisations with two different Elo scales — arena.ai (the former LMArena) and Artificial Analysis. The scores are not comparable across boards. Figures below were read live on July 8, 2026; leaderboards move.

arena.ai (model string dreamina-seedance-2.0-720p)

On text-to-video, Google's gemini-omni-flash leads at 1527 ±13 — though on only 5,449 votes, which arena.ai flags as preliminary, against Seedance's 41,953. Arena reports 1,350,288 total votes across 42 models in the Image-to-Video Arena.

Artificial Analysis (model string "Dreamina Seedance 2.0 720p")

Artificial Analysis states that "Dreamina Seedance 2.0 720p currently leads among Text to Video models with audio output in the Artificial Analysis Text to Video Arena with an Elo score of 1224." On its no-audio text-to-video board, Alibaba-ATH's HappyHorse-1.0 leads at 1290.

The "#1 everywhere" claim

A claim circulating in mid-2026 holds that Chinese labs occupy #1 in every video arena category. On the boards checked here, that is false. Google's gemini-omni-flash is #1 on arena.ai text-to-video. The accurate statement is narrower and still notable: Chinese labs hold #1 on all four Artificial Analysis video boards (Seedance 2.0 on three, HappyHorse-1.0 on the fourth) and on arena.ai image-to-video, and the runners-up in most categories — HappyHorse, Wan, Kling — are also Chinese.

Limitations

• Vendor claims are vendor claims. SeedVideoBench-2.0 is ByteDance's own benchmark, and it publishes charts rather than scores. Every capability claim in the Overview and Capabilities sections is ByteDance's, not an independent finding.
• Unresolved copyright litigation risk. Following viral clips of recognisable actors and characters, The Hollywood Reporter reports that Disney, Paramount, Netflix, Warner Bros. and Sony each sent legal threats to ByteDance in February 2026. On February 20, 2026 the Motion Picture Association sent a cease-and-desist letter; The Hollywood Reporter writes that "it's this first time the MPA has sent a cease-and-desist to a major generative AI company." MPA chairman-CEO Charles Rivkin said in a public statement, quoted by TheWrap: "In a single day, the Chinese AI service Seedance 2.0 has engaged in unauthorized use of U.S. copyrighted works on a massive scale… ByteDance should immediately cease its infringing activity." The MPA letter argued that "the scale and consistency of these results demonstrate systemic infringement rather than inadvertence" (The Hollywood Reporter). Disney's lawyers called the model a "virtual smash-and-grab of Disney's IP" (TechCrunch). ByteDance publicly pledged to strengthen its IP safeguards in mid-February 2026 (CNBC, Variety). The dispute remains unresolved.
• The global rollout was paused because of it. TechCrunch, citing The Information, reported on March 15, 2026 that ByteDance "had planned to make Seedance 2.0 available globally in mid-March" but was "delaying those plans as its engineers and lawyers work to avert further legal issues." Availability outside China arrived later, region by region, and with guardrails — CapCut's initial international release blocked "the ability to make videos from images or videos that contain real faces."
• ByteDance concedes the output is flawed. Its own launch post: "Seedance 2.0 is still far from perfect, with various flaws remaining in its generation results." It names detail stability and hyper-realism as needing refinement, and cites "occasional audio distortion."
• Short clips only. 4–15 seconds. Long-form generation is not supported, and the 30-second capability being marketed belongs to the unreleased 2.5.
• 480p/720p native. No native 4K, and the paper does not claim native 1080p.
• Closed weights. No self-hosting, no fine-tuning, no offline generation, and a dependency on ByteDance's platform terms. For open weights, Wan 2.2 is Apache 2.0.
• Pricing is in CNY and token-denominated. Volcano Engine's rate card is quoted in 元 per million tokens, with the token count a function of resolution, frame rate and duration — so cost is not a simple per-second number, and no USD rate card is published on the Chinese platform.
• Chinese content policy. Generation is shaped by content rules that differ from those governing US-hosted models.

Pricing & Access

Volcano Engine (China) — published in CNY

Prices announced March 4, 2026. Billing is by token count, derived from output resolution, frame rate and duration — not billed directly per second.

Worked example, as reported by ITHome/Sina and BAAI's community hub: a 15-second 720p/24fps clip consumes roughly 308,880 tokens — about 8.65 元 in video-editing mode and about 14.21 元 for pure generation. The latter is a derived figure, and it is the origin of the headline "1 元 1 秒" (one yuan per second) framing; ByteDance does not bill per second. ByteDance publishes no separate rate card for Seedance 2.0 Fast; third-party resellers price it below the standard model, but that is their pricing, not ByteDance's. These are CNY prices and this page does not convert them.

Access routes

• Jimeng / Dreamina — ByteDance's creative platform, one of the launch surfaces for Seedance 2.0 alongside Doubao.
• CapCut (Jianying in China) — integrated into CapCut Video Studio on the paid tier. TechCrunch dates the international rollout to March 26, 2026, beginning in Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand and Vietnam. CapCut's newsroom records expansions to Africa, South America and the Middle East, then Europe (March 28), Japan (March 31) and the United States (April 7, 2026).
• Volcano Engine / Volcano Ark — the China API. At launch the Ark listing read "暂不支持 API 接入" ("API access not yet supported"); it opened later, and on April 14, 2026 Chinese state media reported the API "全面开放" (fully open) to enterprise and individual users.
• BytePlus ModelArk — the international API, whose product page states "Dreamina-seedance-2.0 API Now Fully Available to Clients."

There are no open weights and no self-hosted deployment option. Rates change; confirm against Volcano Engine's model pricing page before budgeting.

Ecosystem & Tools

• Seedance 2.0 product page — ByteDance Seed's official page, SeedVideoBench-2.0 charts, and API links
• Volcano Ark — the China model platform, experience centre and API
• BytePlus ModelArk — the international API surface
• Dreamina — ByteDance's consumer creative platform (Jimeng in China)
• CapCut / Jianying — the video editor Seedance 2.0 is embedded in; The Batch puts CapCut at 736 million monthly active users
• Seedream 5.0 — the sibling image generation line, from the same team and the same Volcano Engine stack
• Seedance 1.5 pro and Seedance 1.0 — the prior generations, both still listed in ByteDance Seed's model directory

Community & Resources

• Official Launch of Seedance 2.0 — the launch post, dated February 12, 2026, and the source of every ByteDance capability quote here
• Seedance 2.0: Advancing Video Generation for World Complexity — the technical report (arXiv 2604.14148, April 15, 2026): resolutions, durations, reference limits, Fast variant
• ByteDance Seed model directory — canonical model list; Seedance 2.0 is the newest video model, and there is no 2.5
• arena.ai Image-to-Video leaderboard — dreamina-seedance-2.0-720p at #1
• Artificial Analysis Text-to-Video leaderboard — #1 with audio
• ByteDance reportedly pauses global launch of Seedance 2.0 — TechCrunch, citing The Information
• MPA Sends Cease and Desist Letter to ByteDance Over Seedance 2.0 — The Hollywood Reporter
• Dreamina Seedance 2.0 comes to CapCut — TechCrunch on the staged international rollout
• Compare with Sora 2, Kling 3.0, Wan 2.2, Hailuo 2.3, Vidu Q3 Pro, and Seedream 5.0