Find any moment in video, with one sentence.
This software is compatible with the smallest model. For higher accuracy, you should use the 3B-parameter model, available at: https://github.com/XinzhengDu/SnapFind_3B
This software is only available for Windows. For Mac or Linux users, please use the Python version:
https://github.com/XinzhengDu/SnapFind_Python
The source code before compilation is available at:
https://github.com/XinzhengDu/SnapFind_Python
Core Logic: All video parsing, model inference, and content retrieval are processed entirely on the user's device.
No user data (videos, images, text) is uploaded to any server. This complies with global privacy regulations:
- China: Personal Information Protection Law (PIPL)
- EU: General Data Protection Regulation (GDPR)
- US: California Consumer Privacy Act (CCPA)
Currently, English query terms are better supported than other languages. If search results are inaccurate, translate your query into English before input.
Before use, you must download the following model and tokenizer files:
- siglip_vision.onnx
- siglip_text.onnx
- tokenizer.json
Download link: https://www.modelscope.cn/models/XinzhengDu/siglip2-base-patch16-224-onnx/files Right-click the file name and select Download.
Place all files in the models folder.
- Content matching: Match search terms with videos, images, and documents. Queries support descriptions of time, location, people, and scenarios.
- OCR matching: Recognize and match text in videos/images (requires more computing power).
- Feature caching: Cache file features to reduce waiting time for subsequent calls.
- Search results: View thumbnails and open source files directly from the results page.
- Frame interval customization: Set N-second frame capture intervals for long/short videos to extract features.
- Precision adjustment: Optimize cache precision for different scenarios to maximize computing resource utilization (Settings > File).
- Low-spec device support: Set longer frame intervals for devices with limited computing power (Settings > File).
- Top-N results: Return the N most relevant results (more results = longer wait time; reduce N for time-sensitive use cases).
- GPU acceleration: 90% complete (supports video loading and model inference via GPU).
- Image upload for search: 90% complete (search for relevant content by uploading images).
- Video player integration: Jump to specific timestamps, add screenshot/GIF features (for meme creation), and audio positioning (for video segmenting).
- Direct file navigation: Open files at specific timestamps (requires designated video player installation).
- Idle-time feature computation: Automatically compute file features when the computer is idle (pauses if resource-intensive programs run).
- Real-time monitoring analysis: Analyze risk levels of live surveillance content (e.g., monitor pet status).
- Vertical model training: Train domain-specific models (requires significant computing power and datasets).
- Locate video timestamps by human actions (e.g., "find the 10th minute where someone waves hands").
- Fast navigation in tutorial videos: Jump directly to relevant sections without watching the entire video.
- Class video indexing: Quickly find key moments in recorded lectures.
- Surveillance footage analysis: Avoid manual frame-by-frame checking (e.g., "find the moment a package is delivered").
- Image-based search: Upload a photo (e.g., a group photo) to find related historical images/videos.
- Legitimate Use Only: This software is strictly for legal and compliant use cases. It is prohibited for illegal surveillance, invasion of privacy, or other unlawful activities.
- User Liability: Users assume full responsibility for all consequences of using this software. The developer shall not be liable for any misuse.
- Content Restrictions: The software restricts retrieval of illegal content (e.g., pornography, violence) to comply with applicable laws and regulations.