Bach-1.0 Preview is an industrial-grade video generation model developed by Video Rebirth, a Singapore-based AI startup founded by former Tencent distinguished scientist Dr. Wei Liu. The model is specifically engineered for professional creative workflows, prioritizing cinematic consistency, high frame rates, and precise directorial control. It produces native 1080p output at 30fps without the need for upscaling or frame interpolation.
The model's architecture is built on two proprietary systems: Dual Diffusion Transformer (DDiT) and Physics-Native Attention (PNA). The DDiT framework is designed to enhance prompt adherence and spatial control, allowing for complex camera maneuvers such as whip pans and rack focus. The PNA mechanism focuses on maintaining temporal continuity and physical realism, ensuring that lighting dynamics and object motion remain consistent across the generated sequence.
One of the primary capabilities of Bach-1.0 Preview is its multi-shot narrative generation. Unlike standard video models that focus on single clips, Bach can generate sequences of up to 30 seconds that include automatic shot planning and transitions. This workflow is supported by a system that locks character identity and environmental details across multiple cuts, addressing the common challenge of character drift in AI-generated films.
Upon its public launch, the model was ranked among the top tier of video generation systems on the Artificial Analysis Video Arena blind benchmark. It supports multimodal inputs, enabling users to combine text prompts with reference images for character identity, location styling, and shot-by-shot direction. The model also integrates secondary generation layers for sound effects and background music within the same production environment.