Volcano Engine Launches Seedance 2.5: Native 30s Video Generation & Enhanced Multi-Character Consistency

火山引擎 · wechat · 2026-08-07

Volcano Engine has officially launched the API for its video generation model, Seedance 2.5. The new version brings significant improvements in instruction following, long-form narrative consistency, cinematic realism, and audio-visual quality. It natively supports 30-second direct video generation, accepts up to 50 multimodal reference materials, and is compatible with over a dozen languages.

Core Upgrades:

Long Takes & Multi-Character Orchestration: Resolves previous issues with character drifting in extended clips. Tests show stable physical traits and size ratios even when 6-7 animal characters cross paths over 30 seconds.
Precise Instruction Following: Highly responsive to second-level timestamps in prompts, allowing for precise execution of cinematic camera movements (e.g., panning, slow-motion) and drastically reducing the "reroll rate" for complex scenes.
Cinematic Realism: Improved rendering of micro-expressions, lighting transitions (e.g., Rembrandt lighting), and physical weight, resulting in a photorealistic feel with a weakened AI look.

Commercial Availability & Pricing:

The model is now available across multiple platforms including LibTV and Qixiang AI. Additionally, older models Seedance 2.0 mini and fast are offering limited-time discounts (Aug 7 - Sep 7), dropping to as low as 0.2 RMB/sec and 0.6 RMB/sec at 720P, respectively.

Original post →

Volcano Engine Launches Seedance 2.5: Native 30s Video Generation & Enhanced Multi-Character Consistency

More from Multimodal