Concepts

Vision-language models (VLMs) combine computer vision and natural language understanding to enable flexible, intuitive interactions with images. Instead of detecting pre-defined object categories, VLMs understand natural language descriptions and questions about visual content.

Datature Vi supports core VLM capabilities that form the foundation of modern vision AI applications.

---

Core VLM capabilities

Dataset types

Choose the right dataset type for your vision AI application:

Explore all dataset types →

Advanced capabilities

---

Which capability should I use?

Choose based on what you need to accomplish:

Learn more:

• Phrase Grounding vs. other vision tasks →
• VQA vs. other vision tasks →
• Freeform for custom use cases →

---

Common use cases

Phrase Grounding applications

Perfect for scenarios requiring object localization with flexible descriptions:

• Robotics — "Pick up the red mug on the left"
• Image editing — Select objects using natural language
• Autonomous vehicles — Identify "the pedestrian crossing from the right"
• Warehouse automation — Find "the damaged box on the top shelf"

Explore detailed use cases and examples →

Visual Question Answering applications

Ideal for extracting information through conversational queries:

• Quality inspection — "Is there a defect on the surface?"
• Accessibility — Describe images for visually impaired users
• Content moderation — "Does this image contain inappropriate content?"
• Inventory management — "How many items are on the shelf?"

Explore detailed use cases and examples →

Freeform applications

🚧 Coming soon — Perfect for specialized scenarios requiring custom annotation formats:

• Research projects — Novel computer vision tasks and experimental approaches
• Medical imaging — Custom diagnostic annotations and measurements
• Scientific imaging — Domain-specific labels and specialized metadata
• Hybrid requirements — Combining multiple annotation types for complex use cases

Explore freeform concepts and use cases →

---

Getting started

Ready to build with VLMs? Start here:

---

Learn more

• Phrase Grounding explained → — Deep dive into visual grounding, how it works, and best practices
• Visual Question Answering explained → — Complete guide to VQA, architectures, and optimization tips
• Chain-of-Thought Reasoning explained → — Learn how step-by-step reasoning improves accuracy for complex visual tasks
• Freeform explained → — Understanding custom annotation schemas and specialized use cases (coming soon)

Related resources

• Phrase grounding — Deep dive into object localization with natural language
• Visual question answering — Understand VQA capabilities and use cases
• Chain-of-Thought reasoning — Learn step-by-step reasoning for complex visual tasks
• Freeform — Custom annotation schemas for specialized use cases (coming soon)
• Quickstart — End-to-end VLM training workflow
• Train a model — Complete training guide
• Annotate data — Create phrase grounding and VQA annotations
• Create a dataset — Set up datasets for training
• Glossary — Common VLM terminology and definitions
• Vi SDK — Python SDK for programmatic access
• Run inference — Use trained models for predictions
• Configure your model — Select model architecture
• Evaluate a model — Assess model performance
• Contact us — Get help from the Datature team