OmniShotCut detects video cuts and transitions at just 164MB

Anyone who has ever broken down video footage knows the pain of scrubbing through a timeline frame by frame just to pinpoint where a cut happened or what transition was used.

OmniShotCut is a new lightweight tool designed to automate that entire tedious process. Feed it a video file, and it will immediately map out every single shot change along with accurate timestamps.

Parsing hard cuts and complex transitions

Shot Boundary Detection has been studied in computer vision for years, but traditional tools usually fall apart the moment editors use creative transitions. They miss subtle visual breaks or generate messy, uninterpretable timestamps when dissolves or zooms blend two scenes together.

OmniShotCut handles both standard hard cuts and complex transition families. It can reliably identify dissolves, fades, slides, and cross zooms, distinguishing clearly between the shots themselves and the transitions connecting them.

Under the hood, the researchers behind the project formulated shot detection as structured relational prediction using a dense Shot-Query Transformer. This allows the system to evaluate both intra-shot and inter-shot relations simultaneously.

Rather than relying on noisy manual annotations, the team built an automated synthetic data pipeline. They gathered 2.5 million raw web videos and generated 300,000 synthetic training clips containing over 11 million parameterized transitions, alongside a modern benchmark called OmniShotCutBench.

Despite its capabilities, the model is remarkably tiny at just 164MB. The code is available for local deployment on regular machines, and the team also launched a free interactive demo on Hugging Face Space.

Why it matters

We think OmniShotCut is a refreshing example of targeted AI engineering that solves a real operational headache.

For creators, editors, and video researchers, having a 164MB model means you can run rapid shot breakdowns completely locally. There is no need to upload confidential footage to expensive cloud APIs or worry about massive GPU clusters; even a basic laptop can crunch through long clips.

For the broader video AI space, this is a vital infrastructure piece. Multimodal models often struggle with long-form video comprehension because standard pre-processing chunks footage arbitrarily, chopping up scenes mid-sentence or mangling camera cuts. Accurate shot boundary detection fixes data ingestion right at the source.

By leaning heavily into high-quality synthetic data instead of bloated parameter counts, OmniShotCut proves that small, hyper-specialized models are often the fastest route to genuine production utility.