Dyence is a platform focused on providing dataset generation services for video and image AI model training, with the core goal of transforming raw video and image materials into AI-ready datasets that can be used directly for training.
In terms of use, Dyence is aimed at teams that need to build training datasets for video or image generation models, helping them transform unlabeled raw media materials into structured, formatted, and verified data.
Its main capabilities include:
- Automatically segmenting videos into clips and identifying scene boundaries. The platform can identify scene changes through mathematical methods on the client side or server side, completing segmentation before calling the API to reduce costs caused by repeated frame analysis.
- Generating multimodal description text covering information such as action descriptions, object categories, speech transcription, and camera position, and can generate rich descriptions including motion, camera angle, and OCR overlay content.
- Providing a Human-In-The-Loop review process, in which professional reviewers manually verify computer-generated timestamps, labels, and object tracking bounding boxes to ensure annotation alignment and clean coordinates.
- Supporting multi-format export, allowing datasets to be exported as structured JSONL lines for easy pushing to Hugging Face, or directly formatted as WebDataset archives.
- Providing interactive dataset previews to visualize the multimodal dataset output results of video and image models.
- Supporting secure cloud archiving, directly compatible with secure cloud object storage architectures, providing fast upload speeds and zero egress fees.
The usage method is divided into four steps: Step one, upload raw videos or images. Supports drag-and-drop folder upload of raw videos or images, or bulk import of links from YouTube or external direct MP4 URLs. Step two, segmentation and annotation. Dyence detects video scene boundaries and outputs rich captions detailing actions, motion, camera angles, and OCR overlays. Step three, human verification. Key training pairs are sent to professional human reviewers to verify caption alignment, correct labels, and clean coordinates. Step four, multi-format export. Export the dataset as structured JSONL lines, ready to be pushed to Hugging Face, or directly formatted as WebDataset archives.
The applicable scenarios are mainly training data preparation for video and image AI models, especially the construction of training datasets for generative models that require high-quality annotation and include multimodal information such as actions, motion, camera, and OCR.
In terms of limitations, the platform uses tiered usage-based pricing and charges only for actual processing volume. The specific prices and output dataset scaleneed to be checked through its price estimator. Details such as the platform's specific processing speed, range of supported video formats, free quota, and subscription plans are not yet provided in the materials.
