OmniShotCut

OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer

OmniShotCut is a sensitive and more informative SoTA for Shot Boundary Detection.
OmniShotCut can detect shot changes of the video in diverse sources (anime, vlog, game, shorts, sports, screen recording, etc.), and recognize Sudden Jump and Transitions (dissolve, fade, wipe, etc.) by proposing a Shot-Query-based Video Transformer.

🔥 Update | 🐍 Quick Start | 🔧 Installation | ⚡ Inference | 🤗 Model Zoo

Update 🔥🔥🔥

• Release ArXiv paper
• Release the inference weights
• Release Gradio demo (with online)
• Release 'pip install omnishotcut' version
• Released the v1.5 weight — a larger model with fewer false cuts on continuous camera motion / lighting changes and better dense fast-cut recall.
• Release the benchmark
• Release the training code and curation

:star: If you like OmniShotCut, please help ⭐⭐star⭐⭐ this repo. Thanks! :hugs:

Quick Start 🐍

First install PyTorch with CUDA support, and make sure a working ffmpeg binary is on your PATH (used for video decoding; e.g. conda install -c conda-forge ffmpeg). Then install OmniShotCut:

pip install git+https://github.com/UVA-Computer-Vision-Lab/OmniShotCut.git

Once installed, running shot boundary detection is just a few lines:

import omnishotcut

# Load model — accepts a local checkpoint path or HuggingFace repo
cut_model = omnishotcut.load("uva-cv-lab/OmniShotCut_v1.5", filename = "OmniShotCut_ckpt.pth")

# Run on a video file
ranges = cut_model.inference("video.mp4", mode="clean_shot")

ranges is a list of [start_frame, end_frame] pairs for each detected shot. By default mode="clean_shot" returns only clean cuts (no transitions). Use mode="default" to also get dissolves, wipes, and fades with their labels:

ranges, intra_labels, inter_labels = cut_model.inference("video.mp4", mode="default")

Besides video file paths, inference() also accepts numpy arrays and torch tensors directly — both (T, H, W, 3) RGB (either uint8, or float in [0, 1]). The input H/W can be arbitrary; frames are resized to the model's process resolution automatically, so results match the video-file path:

ranges = cut_model.inference(frames_thwc, mode="clean_shot")  # frames_thwc: (T, H, W, 3)

Full Local Installation 🔧

conda create -n OmniShotCut python=3.10
conda activate OmniShotCut
conda install -c conda-forge ffmpeg   # required: frame-accurate video decoding backend
pip install -r requirements.txt
pip install -e .

Note: video decoding uses ffmpeg (via ffmpeg-python), so a working ffmpeg binary must be on your PATH. Installing it from conda-forge as above is the most reliable way to get one with all shared libraries present.

Gradio Demo ⚡⚡⚡

Local Gradio can be created by simply running the following:

python app.py 

Click "Running on public URL".

Inference ⚡

This section presents a more formal and controllable way to run inference.

First, let us download the checkpoint

mkdir checkpoints
cd checkpoints
wget https://huggingface.co/uva-cv-lab/OmniShotCut_v1.5/resolve/main/OmniShotCut_ckpt.pth

We provide several inference modes. The 'default' mode outputs the intra and inter labels we define. However, we believe that most users might want the most direct results, which is the general shots without any transitions. To this end, please use '--mode clean_shot'.

Execute the inference by:

python inference.py  --checkpoint_path checkpoints/OmniShotCut_ckpt.pth  --input_video_path __assets__/demo_video1.mp4  --overlap_window_length 10  --mode default

Results are saved to results.json. Visualization is saved to demo_video_results/, where vertical bars with the same color indicate the same shot.

Model Zoo 🤗

Model	Description	Huggingface
OmniShotCut v1.5 (recommended)	Hidden dim 576, 52.8M params. Best real-world quality: fewer false cuts on continuous camera motion / lighting changes and better dense fast-cut recall.	Download
OmniShotCut (paper)	3 encoders, 34.5M params. Reproduces the paper (Table 1) results.	Download
OmniShotCut (preview)	6 encoders, 41M params. Early preview release.	Download

v1.5 is a drop-in replacement (same filename and interface): it has fewer false cuts on continuous camera motion and lighting/strobe changes, recovers more dense fast cuts, and keeps the same accuracy on clean content.

Awesome work using OmniShotCut

• Hy-Nomas: Uses OmniShotCut for shot-boundary detection to split raw videos into clips during data preprocessing for interactive world generation.
• DreamX-Creator: Applies OmniShotCut for fine-grained shot-boundary detection and transition labeling during finetuning-data curation, discarding candidate segments that contain or span a detected transition.
• Sekai2: Uses OmniShotCut to detect hard cuts and gradual transitions, then partitions continuous perspective shots into clean clips (up to 120s) for world-model pretraining data.
• ContextMaster: Invokes OmniShotCut to segment multi-shot videos and localize individual shots, both for its Interactive Director tool and for preparing addition/removal/replacement training data.
• AlayaWorld: Combines classical cut/dissolve detection with OmniShotCut in its data-curation pipeline's shot-boundary filtering stage, ensuring every retained training sample is a single continuous shot.

📚 Citation

@article{wang2026omnishotcut,
  title={OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer},
  author={Wang, Boyang and Xu, Guangyi and Zhang, Jiahui and Tang, Zhipeng and Cheng, Zezhou},
  journal={arXiv preprint arXiv:2604.24762},
  year={2026}
}