Skip to Content

12. Multimodal AI

Most production AI systems in 2026 handle more than text. If you’ve only built text-in-text-out, this module fills the gap.

Vision — Image Understanding

Send an image to the model alongside a text prompt. GPT-4o, Claude, and Gemini all accept images natively. Use cases: document extraction (receipts, forms, diagrams), visual QA, image classification without training a custom model. Start by sending a screenshot and asking the model to describe it — then try structured extraction.

Image Generation

DALL-E 3 (via API), Stable Diffusion (open-source, run locally), and Midjourney (no API). For programmatic use: DALL-E API or Stability AI’s API. Understand that generation is a separate model from understanding — the same system prompt patterns don’t apply.

Audio — Speech-to-Text and Text-to-Speech

OpenAI Whisper (API or local) is the standard for transcription. Handle long audio by chunking into segments. Text-to-speech: OpenAI TTS API, ElevenLabs, or open-source Bark/Coqui.

!!! warning “Voice cloning” Voice cloning raises ethical and legal questions — know the boundaries.

Video Understanding

Extract frames, send as image sequence with timestamps. Gemini handles long video natively. For others, sample keyframes and process as images. Video is expensive in tokens — be selective about what you send.

Multimodal RAG

Embed images alongside text using multimodal embedding models (Cohere Multimodal, Jina CLIP). Store in the same vector DB. Retrieve across modalities — a text query can surface relevant images, and vice versa.

Resources

OpenAI Vision · OpenAI Whisper · OpenAI TTS · DALL-E API · Gemini multimodal · Stable Diffusion · Hugging Face multimodal models