General · EdgepediaSeptember 17, 20266 min read

Seedance

Seedance is a family of text-to-video and multimodal audio-video generation models developed by ByteDance's Seed team, first released in 2025 and expanded through the 2.0 and 2.5 generations in 2026. The series began as a high-speed video foundation model and evolved into a unified system that generates picture and sound together, accepts text, image, audio, and video inputs as references, and supports multi-shot storytelling and timestamp-level editing.¹^•²

Key fact	Detail
Developer	ByteDance Seed team (Beijing-based ByteDance, creator of TikTok)³
First release	Seedance 1.0, a video foundation model integrated into Doubao and Jimeng in June 2025¹
Speed	Seedance 1.0 generates a 5-second 1080p video in 41.4 seconds, aided by a 10x inference speedup from multi-stage distillation¹
Audio-video generation	Seedance 1.5 pro introduced native joint audio-video generation with multilingual and dialect lip-syncing⁴
Seedance 2.0	Released in China in early February 2026; 4 to 15 second audio-video outputs at 480p and 720p; up to 3 video clips, 9 images, and 3 audio clips as references²
Seedance 2.5	Launched July 31, 2026; 30-second single-pass generations, up to 30 images, 10 videos, and 10 audio references, and timestamp-level editing⁵
Controversy	The Motion Picture Association sent ByteDance a cease-and-desist letter in February 2026, its first to a major AI firm, alleging infringement of member studios' works⁶

Seedance 1.0

Seedance 1.0 was introduced as a high-performance, inference-efficient video foundation generation model. Its technical report describes four core improvements: multi-source data curation with precision video captioning; an efficient pre-training paradigm enabling interleaved multimodal positional encoding, native multi-shot generation, and multi-task modeling; post-training with fine-grained supervised fine-tuning and video-specific RLHF using multi-dimensional reward mechanisms; and model acceleration achieving a 10x inference speedup through multi-stage distillation and system-level optimizations.¹

The model supports multi-shot video generation from both text and image, producing 1080p videos with consistent subjects, visual style, and atmosphere across shot transitions. On the third-party benchmarking platform Artificial Analysis, Seedance 1.0 ranked first on both the text-to-video and image-to-video leaderboards as of June 2025.⁷ It was integrated into ByteDance's Doubao and Jimeng platforms in June 2025.¹

Seedance 1.5 pro

Seedance 1.5 pro, described in an arXiv report, was engineered specifically for native, joint audio-video generation. It uses a dual-branch Diffusion Transformer architecture with a cross-modal joint module and a multi-stage data pipeline, plus post-training with supervised fine-tuning and RLHF, and an acceleration framework that boosts inference speed by over 10x. ByteDance highlights precise multilingual and dialect lip-syncing, dynamic cinematic camera control, and enhanced narrative coherence; the model is accessible on Volcano Engine.⁴ The 1.5 pro version added joint audio-video generation with lip-synced multilingual dialogue and sound effects, launching at 480p and 720p.⁸

Seedance 2.0

Seedance 2.0 was officially released in China in early February 2026, with the launch announced on February 12, 2026.²^•⁹ It adopts a unified multimodal audio-video joint generation architecture supporting four input modalities: text, image, audio, and video.⁹ Users can simultaneously input up to 9 images, 3 video clips, and 3 audio clips plus natural-language instructions, and the model can reference composition, motion, camera movement, visual effects, and audio from those inputs, including text-based storyboards.⁹

The model generates audio-video content of 4 to 15 seconds at native 480p and 720p resolutions, supports 15-second multi-shot output with dual-channel audio, and ships with an accelerated variant, Seedance 2.0 Fast, for low-latency scenarios.²^•⁹ ByteDance's own evaluation reports industry-leading performance in multimodal reference generation and video tasks, while acknowledging remaining weaknesses in detail stability, hyper-realism, occasional audio distortion, multi-subject consistency, and text rendering accuracy.⁹

Seedance 2.0 rolled out on Jimeng AI, Doubao, and other platforms, with API access via BytePlus ModelArk and resale through third-party hosts such as fal.ai, Replicate, WaveSpeed, and Kie.¹⁰^•⁸

Seedance 2.5

Seedance 2.5 launched on July 31, 2026, building on the 2.0 architecture with emphasis on long-form storytelling, multimodal reference, and editing. It extends single-pass generation from 15 to 30 seconds with multi-round extensions, allowing multi-minute content with a consistent audiovisual language, and raises reference limits to 30 images, 10 video clips, and 10 audio clips in a single pass.⁵ Editing gains include timestamp-level control for targeted changes to audio and video, plus green screen, camera perspective, and reference-based editing; the model also uses clay-render references to control camera movement, blocking, and physically consistent lighting. It rolled out on Jimeng AI and Doubao Pro, with API access via BytePlus ModelArk to follow.⁵

Hollywood copyright dispute

Shortly after launch, Seedance 2.0 drew condemnation from Hollywood organizations. A widely shared video posted by Irish director Ruairí Robinson showed AI versions of Tom Cruise and Brad Pitt fighting in a post-apocalyptic wasteland; screenwriter Rhett Reese, co-writer of the Deadpool movies, responded that the technology threatened film careers.³ The Motion Picture Association said Seedance 2.0 had "engaged in unauthorized use of U.S. copyrighted works on a massive scale," and the actors union SAG-AFTRA condemned what it called unauthorized use of members' voices and likenesses.³

On February 20, 2026, the MPA sent ByteDance a cease-and-desist letter, the first time the association, which represents all of the major Hollywood studios, had sent such a letter to a major AI firm. The letter, addressed to ByteDance global general counsel John Rogovin, accused the company of producing content that closely replicates distinctive characters and other copyright-protected elements of member studios' works, and demanded written confirmation of corrective steps by February 27. In the following days, Disney, Warner Bros. Discovery, Paramount Skydance, Netflix, and Sony Pictures each sent their own cease-and-desist letters.⁶ ByteDance responded that it respects intellectual property rights and is taking steps to strengthen safeguards against unauthorized use of intellectual property and likeness by users.³ At the time of the backlash, Seedance 2.0 was available only in China.³

The dispute intersected with commercial concerns about cost. In testimony before the United States Senate, one content creator shared a comparison of a clip from the film F1 with a near-exact copy generated by Seedance, claiming the model replicated the most expensive shot in the film for nine cents.¹¹

References

Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Model families and named models › Video generation models

Initially written Sep 17, 2026 · Reviewed: Sep 18, 2026; Oct 15, 2026 · Edited: Sep 18, 2026; Oct 15, 2026 · Last review: Oct 15, 2026

Notice something wrong?