Chapter 5: Multimodal Prompting#
Why it matters#
Every technique so far has worked on text. But the world is not only text. People perceive through many senses at once and communicate with gestures, expressions, and images as much as words. Multimodal models bring that richness to generative AI, accepting and relating multiple data types, especially text and images, so a single model can look at a picture and talk about it. This closing chapter of Module 1 explains why multimodality matters, what a multimodal LLM is, how to prompt one, and the use cases it unlocks. It also sets up Module 3, where multimodal retrieval and agents appear.
A motivating problem#
Imagine a marketing company hired by a pet shop. The shop needs an individual flyer for each pet, generated from the pet’s image and a short bio. How would you do this with traditional, single-modality models?
An image-only model (computer vision) learns only from images. It can classify images, detect objects, and segment scenes, so it could sort pets into “cat” and “dog” classes. But it cannot read the bio text or generate the flyer’s written description.
A text-only model learns only from text. It can generate a campaign description from a written brief, but it cannot see the pet, so it cannot describe an animal from its photo.
Neither model alone solves the task. You need one system that understands both the image and the text, and that is precisely what a multimodal model provides.
What “multimodal” means#
Humans are naturally multimodal. We perceive the world through vision, hearing, smell, taste, and touch, and we communicate non-verbally through gestures, facial expressions, body language, eye contact, and appearance. Multimodality in AI is an attempt to give models a similarly broad channel to the world.
The motivation is practical: generative AI has shifted from prediction to interaction, and handling multiple modalities is a direct way to make AI better at interacting with people to solve real problems.
Data modalities#
This course focuses mainly on two modalities, and it is worth understanding why.
Modality |
Why it matters |
|---|---|
Image |
The most versatile input format. There is far more visual data than text data, phones and webcams generate it constantly, and images can represent text, tabular data, audio (as spectrograms), and to some extent video. |
Text |
A powerful output format. A model that understands and generates text can summarize, translate, reason, and answer questions. |
Other modalities, video, audio, haptic data, and electrical signals, exist too, but text and image are the focus here.
Multimodal LLMs#
Recall that an LLM is a foundation model trained on text whose core skill is predicting words in context. A multimodal LLM (MLLM) is a large language model trained on multiple modalities.
The common architectural trick is to keep a capable language model at the core and equip it with cross-modal capabilities using encoders and adapters. For example:
a vision encoder converts an image into a representation the language model can attend to,
a video encoder does the same for video,
an audio encoder does the same for sound.
In other words, the encoders translate non-text inputs into the same kind of internal representation, embeddings, that the language model already knows how to work with. This is the same embedding idea introduced with Titan Embeddings in Chapter 1, now applied across modalities.
Prompting multimodal LLMs#
Prompting an MLLM combines everything from Chapter 3 with a new set of considerations for images.
Text prompts follow the same best practices as before: clear, specific, positive, well-structured instructions.
Image prompts add format and quality constraints that you must respect or the model will reject or misread the input:
Consideration |
Guidance |
|---|---|
Input format |
Most MLLMs expect base64-encoded images. |
Image size |
Stay within size limits (for example, under ~5 MB). |
Multiple images |
Most MLLMs can analyze only a limited number of images per request. |
Image format |
Use a supported format (jpg, png, etc.). |
Image clarity |
Avoid blurry images. |
Image placement |
It often works better when the image comes before the text. |
Image resolution |
Stay within the model’s resolution limits. |
Worked example: the pet-shop flyer, revisited
With an MLLM, the pet-shop task becomes a single prompt: supply the pet’s photo (base64-encoded, placed first) followed by a text instruction such as “Using the image and this bio, write a warm one-paragraph adoption flyer highlighting the pet’s appearance and personality,” then the bio text. One model now reads the image and the bio and writes the description, the task that defeated both single-modality models.
Multimodal use cases#
Several application patterns recur across industries.
Visual question answering (VQA). Give the model both text and an image and ask about the image, for example, “What is the purpose of the highlighted part in this circuit board?” The model generates a text description or answers a question grounded in the picture.
Text-based image retrieval. Given a text query, find the images whose captions, metadata, or embeddings are closest to the query, “find chairs in stock,” returning matching product images. This matters not only for search engines but for enterprises searching their internal images and documents.
Deep image similarity retrieval. Given an image, find similar images, for example retrieving visually similar Amazon products or identifying other products from the same manufacturer.
AWS in practice
On Amazon Bedrock, multimodal capability shows up in two complementary places. Multimodal generative models accept images alongside text for tasks like visual question answering and captioning. Multimodal embedding models (in the Titan family) turn images and text into a shared vector space, which is exactly what powers the retrieval use cases above and the retrieval-augmented generation you will build in Module 3.
In the news#
Multimodality has become a baseline expectation rather than a special feature. Leading models now routinely accept images alongside text, and the modality set keeps widening toward audio and video, moving the field closer to the multi-sense interaction this chapter described. On AWS, the Titan and newer Nova model families provide both multimodal understanding and multimodal embeddings, bringing the patterns here, visual Q&A, image retrieval, and similarity search, into managed services. As always, consult the Amazon Bedrock documentation for the current multimodal model lineup.
Hands-on labs#
Bring the module to a close with Lab 5: Multimodal Prompting, which prompts a multimodal model on Amazon Bedrock using both images and text and explores visual use cases hands-on.
Key takeaways#
Single-modality models cannot solve tasks that require both seeing and writing; multimodal models can.
An MLLM keeps a language model at its core and adds encoders/adapters to bring images, video, and audio into a shared representation.
Prompting an MLLM combines text best practices with image constraints (base64 encoding, size and format limits, clarity, and placement).
Key use cases are visual question answering, text-based image retrieval, and image similarity retrieval, all powered by embeddings.
This completes Module 1. You now understand foundation models and LLMs, the transformer architecture, prompt engineering from basic to advanced, and multimodal models, the full toolkit for working with foundation models on Amazon Bedrock. Module 2 turns to using these models responsibly: evaluating them, applying responsible-AI principles, and improving their security and safety.