Seedance 2.0: ByteDance's Next-Gen AI Video Generator

Seedance 2.0 entered beta on February 7, 2026, rolling out first in CapCut for Chinese users. The global release date is expected around mid-February 2026, with availability through Editly,Dreamina以及其他平台。字節跳動的 Seed 團隊在原定時間表之外的幾個月裡一直在對模型進行優化。這個版本在 Seedance 1.0 Pro（2025 年 6 月）和 1.5 Pro（2025 年 12 月）所提供的基礎上，新增原生多鏡頭敘事、2K 輸出，以及更緊密的影音同步。

Version History

• Seedance 1.0 Pro, June 11, 2025. Text-to-video and image-to-video at 1080p/24fps, 5-10 second clips. Topped both T2V and I2V leaderboards at launch. Free access on Dreamina.
• Seedance 1.5 Pro, December 16, 2025. Added joint audio-visual generation in one pass, with millisecond-level lip sync across six languages. Cinematic camera control and 10x inference speedup through distillation.
• Seedance 2.0, beta on February 7, 2026 (CapCut, China). Global rollout expected mid-February. Originally planned for late 2025, delayed for further optimization. Beta testers describe the output as noticeably better than 1.5 Pro.

The Backstory

Before Google dropped Veo 3, the Seedance team was still searching for direction. Veo 3 showed them what the market actually needed. They scrapped the old roadmap and rebuilt from scratch.

That pivot produced Seedance 1.5 Pro, which was already solid. The team then had 2.0 in a shippable state around November-December 2025, but felt the quality could go further. They spent another two to three months polishing it.

Beta testers with early access say the gap between 1.5 Pro and 2.0 is hard to miss. Chinese users on the CapCut beta have been generating short films and ads since launch day. One tester produced a Shaw Brothers-style martial arts short with synchronized audio and free-moving camera in a single prompt. Another generated a 15-second Demon Slayer-style animation from text alone. The delay looks deliberate: ByteDance is taking its time to get this right.

What's New

Multi-shot storytelling. A single prompt generates multiple coherent shots with consistent characters, maintaining the same face and outfit across camera angles.

2K resolution, 30% faster. Up from 1080p. A 2K clip generates in about 60 seconds. Supports six aspect ratios (16:9, 9:16, 4:3, 3:4, 21:9, 1:1). Clips run 5-15 seconds.

Multimodal input with up to 12 reference files. Mix images, videos, and audio. Use an image to lock the visual style, a video for motion and camera movement.

Audio generated in one pass. A Dual-Branch Diffusion Transformer handles dialogue, Foley, and ambient sound alongside video. Phoneme-level lip sync in 8+ languages.

In-video editing. Swap characters, add or remove objects, extend clips, canvas expansion, and in-painting, all through text instructions.

比較方式

Sora 2 在物理模擬精確度與單鏡頭寫實方面更佳，但原生音訊較弱，且不支援多鏡頭敘事。

Veo 3.1 是目前的逼真度基準，在一個包含 1,003 個提示的評估中得分最高。它也具備原生音訊，但成本較高且較不易取得。

Kling 2.6 提供強大的原生音頻生成能力，以及一個可從參考視頻複製動作的動作轉移功能。解析度上限較低。

Seedance 2.0 在多鏡頭的一致性與多模態參考輸入方面具備差異化特徵，並得益於字節跳動的速度優勢。

在哪裡存取

CapCut 的測試版現已面向中國用戶上線。全球可用性預計於二月中旬起：

• Editly, 整合正在進行中，模型在全球可用後將立即上線。
• Dreamina (dreamina.capcut.com), 主要的第一方平台
• 像 Replicate 與 Segmind 這樣的第三方 API

Dreamina 一直以來為新版本的 Seedance 提供免費階層的存取權。

結論

多重敘事生成與12檔案的多模態輸入是本次的重點特徵。Beta 測試者的反饋為正向。一旦模型公開，獨立基準測試將補充細節。