Seedance 2.0 by ByteDance: status and timeline

Seedance 2.0

Multimodal system synthesizing audio and video simultaneously. Takes text, images, video, and audio. Outputs 4-15 seconds at native 480p-720p resolution.

DeployStage 4 of 5

Released February 2026; ranks #2 on Artificial Analysis (Elo 1,269) for video generation as of March 2026.

Updated 12 Feb 2026Checked 10 Oct0 updates this week

Milestones

No announced next step

Seedance 1 releasedJun 2025Complete.
Seedance 2.0 announced12 Feb 2026Complete.

Most important updates

Seedance 2.0 released with unified audio-video generation12 Feb 2026

Current obstacles

Long-duration coherenceGenerating longer sequences without drift in object identity or physics requires better long-context understanding.

Physics limits

Compute cost per frameVideo generation requires synthesizing dozens of high-resolution frames per second. Each second of video means far more compute than a text answer.
Training data distributionModels learn patterns from training data. Rare events or novel object combinations are harder to generate reliably than common scenarios.

How it works

3 parts

Architecture

Dual-branch diffusion transformer

Seedance uses separate pathways for video and audio latents, connected via cross-attention. This unified approach generates coherent synchronized output.

Inputs

Rich multimodal composition

Users can provide text instructions along with reference images, video clips, and audio clips. The model synthesizes these into a single coherent video.

Physics modeling

Neural Motion Transformer learns real-world physics

The model implicitly learns gravity, inertia, and light refraction, applying them frame-by-frame. This reduces artifacts in object interactions and motion.

Spec sheet

Spec	Seedance 2.0
Video duration (native)	4-15 seconds^R (reported)
Native resolution	480p-720p (2560×1440 upscaled)^R (reported)
AA Elo rating (March 2026)	1,269^R (reported)
Generation speed vs Seedance 1	30% faster^R (reported)
Phoneme-accurate lip-sync languages	8+ languages including dialects^R (reported)

^R reported by the company

Update log

1 update

Thu 12 Feb

Major: Seedance 2.0 released with unified audio-video generationPressSoftware

About ByteDance

The team behind Seedance 2.0

ByteDance

TikTok, Doubao, Seedance, CapCut, Douyin

Beijing-based platform company behind TikTok and Douyin with 1B+ users each. AI division produces Doubao (conversational model) and Seedance (video generation). Exceeded Meta on quarterly revenue in Q2 2025.

Doubao reached 382M monthly active users by June 2026, dominating China's consumer AI market.
Seedance 2.0 generates video and audio simultaneously; ranked #2 on Artificial Analysis (Elo 1,269) as of March 2026.
AI models drive content recommendation and generation; core to ByteDance's $120B+ first-half 2026 revenue.

bytedance.com (opens bytedance.com)Company profile →

Founded: 201214 yrs
Headquarters: China
Status: Private
Valuation: $550BprivateFeb 2026
Staff: ~110,000Jan 2026
Works in: AIFrontier models
Coverage: 2 programs · 3 updateslatest 1 Jun 2026checked 10 Oct
People: Zhang Yiming (opens en.wikipedia.org)FounderLiang Rubo (opens en.wikipedia.org)CEO

Spec

Seedance 2.0

Video duration (native)

4-15 seconds^R (reported)

Native resolution

480p-720p (2560×1440 upscaled)^R (reported)

AA Elo rating (March 2026)

1,269^R (reported)

Generation speed vs Seedance 1

30% faster^R (reported)

Phoneme-accurate lip-sync languages

8+ languages including dialects^R (reported)