Skip to content

Multimodal AI: vision, audio, video

How models take in pixels, audio and frames — what that costs, where it fails, and which parts of the job still belong to a specialised tool.

A multimodal model is not a language model with a camera bolted on. An image is turned into tokens, those tokens go through the same transformer as your text, and almost everything that follows — the price, the resolution limit, the peculiar failures — follows from how that conversion works.

These pages start from the conversion and work outward. Where a classical tool still wins, they say so: OCR engines, detectors and speech recognisers did not stop being good at their jobs because a general model learned to do them adequately.

How a Multimodal Model Sees an Image

An image becomes a sequence of tokens and then goes through an ordinary transformer. Traced end to end, with the patch arithmetic.

6 min read

Image Tokens: How Pictures Are Priced

The published tiling formulas, worked through on real image dimensions, and the resizing you should do before the provider does it for you.

6 min read

Detail Levels: When Low-Res Image Input Is Enough

What the detail setting actually changes, a minimum-feature-size rule for choosing it, and how to A/B it on your own images.

6 min read

OCR With a Vision Model vs a Real OCR Engine

The two systems return different things, fail differently, and price differently. Which differences are structural, and how to decide.

6 min read

Document Understanding: PDFs, Tables and Layout

Where table extraction still breaks, shown as concrete structures, and what to do about each one.

7 min read

Chart and Diagram Reading: What Models Get Wrong

Why reading a chart is harder than reading a table, which encodings break, and what the published chart benchmarks are actually testing.

6 min read

Video Understanding: Frames, Sampling and Cost

Video is a stack of images with a token bill attached. Frame sampling is the main lever, and it is a cost decision before it is a quality one.

7 min read

Audio Input: Speech-to-Text vs Native Audio Models

A transcription pipeline and a native audio model differ in what information survives, not only in latency. Which difference matters for which job.

7 min read

Text-to-Speech in 2026: Latency, Voices and Cost

What actually determines time-to-first-audio, how to measure it yourself, and the axes other than latency that decide a TTS choice.

6 min read

Real-Time Voice Agents: The Full Latency Budget

Every term between the microphone and the speaker, what sets each one, and which ones are actually worth attacking.

7 min read

Image Generation Models: A Practical Comparison

The structural axes that actually differ between image model families, and a protocol for running the comparison on your own prompts.

7 min read

Prompting for Images vs Prompting for Text

Why skill at prompting language models transfers badly to image models, traced to the difference in how each one reads your words.

6 min read

Image Editing and Inpainting via API

Mask-based editing, instruction editing, and the production details — alpha channels, feathering, seams and re-encoding — that decide whether it holds up.

7 min read

Multimodal RAG: Retrieving Over Images and Text

Joint embedding versus caption-then-index versus retrieving over page images — what each costs in storage, and what each fails at.

7 min read

Vision Model Hallucinations: Seeing Things That Aren't There

A taxonomy of the ways a vision model reports things the image does not contain, anchored to the published literature, with detection strategies for each.

7 min read

Screenshot Understanding for UI Automation

Getting a model to click the right pixel: the coordinate-space transform that causes most of the error, and how to test your own accuracy.

7 min read

Handwriting and Low-Quality Scans

How a general vision model and a dedicated handwriting engine differ on messy input, and how to measure which one is better on yours.

6 min read

Bounding Boxes and Grounding: Getting Coordinates Out

Which approaches return coordinates at all, in what coordinate space, and how to measure whether they are accurate enough for your use.

6 min read

Image Safety Filters: Why Your Benign Request Was Blocked

Where in the stack a block happens, the categories of benign request that trip classifiers, and how to build a product that copes with refusals.

6 min read

Building a Multimodal Pipeline That Doesn't Cost a Fortune

Downscale, cache and cascade — the three levers that move a multimodal bill, with the arithmetic for deciding whether each is worth it.

7 min read

Multimodal AI: vision, audio, video · Multigrid