On May 7, 2026, Video Rebirth launched BACH 1.0, a new AI video generation engine aimed at multi-shot, production-grade output. The release matters less for the headline benchmark and more for what BACH is trying to fix: the gap between generating a clip and generating a film.
Here is what BACH does, what is genuinely new, and where it sits in the model landscape today.
What BACH 1.0 Actually Is
BACH is built by Video Rebirth, a Hong Kong-based company founded by Dr. Wei Liu, formerly a distinguished scientist at Tencent. The company raised $80 million in earlier funding to build what it calls an “industrial-grade” AI video engine.
The launch ships five creation modes:
- Montage — multi-shot sequences up to 30 seconds with consistent characters
- Text to Video — a single clip from a text prompt
- Image to Video — animate a still image
- Element to Video — generate video from a specified visual element
- Create Image with one-click video conversion — image generation feeding directly into a clip
Output is native 1080p at 30fps — not upscaled or interpolated. Sound design is generated alongside the visuals.
Why Multi-Shot Output Is the Interesting Part
Most current AI video models generate a single clip per call. To make anything longer than a few seconds, you stitch clips together — which means losing character identity, lighting continuity, and camera language at every cut.
BACH’s Montage mode generates multiple shots inside one workflow. The promise is consistent characters and coherent transitions across the full 30-second output without the manual chaining step.
That is the same problem Tellers’ agent solves at the workflow level — orchestrating multiple models, references, and edits to produce a final timeline. A model that handles part of it in a single pass is useful infrastructure for that kind of agent.
The Architecture in One Paragraph
BACH is built on two proprietary architectures. Dual Diffusion Transformer (DDiT) is the part that translates cinematographic instructions — camera moves, depth of field, lens choices, lighting — into physically accurate motion. Physics-Native Attention (PNA) handles character identity using bone structure and muscular dynamics rather than visual similarity alone, which is meant to keep characters recognizable across shots and motion.
The technical claims are not independently verified yet. What is verified is the benchmark: BACH 1.0 Preview ranked #6 globally on the Artificial Analysis Video Arena before this public release.
Where BACH Sits in the Model Landscape
The shortlist of competitive AI video editing and generation models in May 2026 looks roughly like this:
- Runway Gen-4.5 — strong on physics and motion control
- Kling 3.0 — native 4K with audio in one pipeline
- Veo 3.1 — strong text adherence; Veo 4 expected at Google I/O later this month
- Hailuo 2.3 — character micro-expressions and stylization
- Seedance 2 — ByteDance’s general-purpose video model
- HappyHorse-1.0 — currently #1 on Artificial Analysis Video Arena
BACH enters at #6 on the leaderboard with a different bet: that multi-shot, character-consistent output in a single workflow matters more than another marginal jump on single-clip quality. Whether that bet pays off depends on how well the consistency holds up under real production loads — not on the leaderboard.
What This Means for Tellers Users
We have not integrated BACH yet. As a multi-model platform, Tellers evaluates new video models against the workflows our users actually run. Leaderboard rank is one signal among several.
The Tellers agent already orchestrates AI video creation across Runway Gen 4.5, LTX Video, Kling, Hailuo, Seedance 2, Veo 3.1, and others. Multi-shot consistency is currently handled at the workflow layer — references, entities, prompts carried across calls — rather than inside a single model. If BACH’s Montage mode delivers on character consistency, it could shift where that work happens.
We will share an update if and when we integrate it.
FAQ
What is BACH 1.0?
BACH 1.0 is an AI video engine launched on May 7, 2026 by Video Rebirth, a company founded by former Tencent distinguished scientist Dr. Wei Liu. It generates native 1080p, 30fps video, with a multi-shot mode called Montage that produces up to 30 seconds of coherent footage in a single workflow.
What makes BACH different from other AI video models?
Two things. First, its Montage mode generates multi-shot sequences with consistent characters and camera language in one pass, rather than chaining individual clips. Second, the model is built on two proprietary architectures — Dual Diffusion Transformer (DDiT) for cinematographic control and Physics-Native Attention (PNA) for identity and physical consistency across frames.
Where does BACH rank against other AI video models?
The BACH 1.0 Preview ranked #6 globally on the Artificial Analysis Video Arena before this public release. The launch version is positioned as a direct competitor to Veo, Runway Gen-4.5, Kling, Hailuo, and Seedance.
Is BACH 1.0 available on Tellers?
Not yet. We will evaluate the public BACH API and integrate it if it adds meaningful capability to the Tellers agent. Tellers currently supports Runway Gen 4.5, LTX Video, Kling, Hailuo, Seedance 2, Veo 3.1, and others. The model list is visible inside the app.
Who is BACH 1.0 built for?
Video Rebirth targets e-commerce, film and television studios, short-form drama, advertising agencies, and game studios. The Montage mode in particular is aimed at production pipelines that need multi-shot output rather than single clips.
Is BACH 1.0 open source?
No. The model is available via web product at bach.art with free credits on signup. Commercial use requires a paid plan. There is no public release of weights or code.
New AI video models will keep arriving. The interesting question is not which one ranks first this week, but which ones meaningfully change what an editor — human or agent — can do. If you want a workflow that already combines multiple top models with full timeline editing, open Tellers and try it on a real project.