Visual Layer’s Enrichment Hub lets you generate high-value metadata using pre-trained models tailored for image and video datasets. These enrichment models can:

  • Extract descriptive labels, captions, and tags
  • Enable advanced semantic and object-level search
  • Power downstream filtering, QA, and automation
  • Improve annotation coverage and data understanding

You can apply one or more models during the enrichment process, depending on the type of insights you want to generate.

​
Available Enrichment Models

Visual Layer provides a wide range of built-in models designed for diverse enrichment tasks:

Model NameTask TypeDescription
VL-Object-DetectorObject DetectionIdentifies and localizes objects with bounding boxes and class labels.
VL-Image-TaggerMulti-Class ClassificationApplies multiple labels to the entire image for categorization and metadata generation.
VL-Object-CaptionerObject to TextGenerates short captions describing individual objects in context.
VL-Image-CaptionerImage to TextSummarizes the scene or image with natural language.
VL-Image-Semantic SearchSemantic Image SearchEnables conceptual search over images using natural language queries.
VL-Object-Semantic SearchSemantic Object SearchEnables contextual search for specific objects based on semantics.
NVILA-Lite-2BImage-Text-to-Text (VQA)Efficient VQA model for visual understanding tasks across multiple frames or images.
Janus-Pro-1BImage-Text-to-Text (VQA)Autoregressive model for multi-modal reasoning and question answering.

Important

Some models require pre-existing labels before enrichment.
These models include:

  • Object/Image Captioners
  • Semantic Search (Object or Image)

Labels may come from user annotations, the Object Detector, or the Image Tagger model.

​
Coming Soon

These models are in development and will be available soon in the enrichment catalog:

  • GPT-4o – Multimodal LLM for advanced captioning and reasoning
  • Nv-grounding dino – Grounded object detection with natural language prompts
  • Neva-22B – High-performance image-to-text transformer
  • Nvclip – Lightweight vision-language model for fast embeddings
  • Qwen-VL-2B – VQA and captioning model for diverse use cases
  • Moondream2 – Compact vision-language model
  • Molmo-7B – Multimodal model for contextual enrichment
  • Advanced-Image-Search – Conceptual image retrieval using complex queries
  • Advanced-Object-Search – Enhanced semantic object retrieval
  • VL-Object-Tagger – Assigns tags at the object level across frames or images

​
Want Early Access?