Images, photos, and multimodal basics

Multimodal systems can combine text with an image, audio clip, or document. This can make assistance more flexible, but errors can occur in every format.

A real photograph still has a viewpoint, a date, and a context. Cropping, captions, and edits can change how people interpret it.

To a computer, a photograph is nothing but numbers — a grid where every pixel is recorded as three values for red, green, and blue. Computer vision is the task of turning that grid of numbers into a useful statement such as 'this leaf shows signs of maize streak virus,' and the way modern systems do it is one of the clearest illustrations of learning from examples. The workhorse is a neural network organised in layers. The first layers examine tiny neighbourhoods of pixels and learn to respond to elementary patterns: an edge between light and dark, a patch of a particular colour, a simple texture. The next layers combine those responses into larger structures — a corner, a vein pattern, a curved outline. Deeper layers combine structures into parts, and the final layers combine parts into a judgement, producing a score for each possible label. Nothing at any layer was programmed by hand; during training the network saw many thousands of labelled photographs and repeatedly adjusted its internal numbers so that its scores matched the labels more often. Follow a worked example: a crop-health app is trained on twenty thousand photos of maize leaves, each labelled healthy or diseased by agronomists. Early layers learn to detect the yellow streaking pattern as a texture; middle layers learn where on a leaf it typically appears; final layers learn how much streaking separates mild from severe cases. Point the camera at a new leaf and the network runs the same arithmetic forward and reports, for instance, diseased with 87 percent confidence. That number measures similarity to the training examples, not certainty about the world — a leaf photographed at dusk, or a maize variety absent from the training set, can produce a confident and wrong answer, because the pixel patterns fall outside everything the network learned. The misconception to correct is that the system sees the way people do, understanding that it is looking at a plant in a field. It maps pixel statistics to labels, nothing more, which is why a sticker on a road sign can fool a vision system that no human would misread. Practically, this means treating a vision result as a fast, tireless first opinion — excellent for screening many images — while keeping a person responsible whenever the answer matters.

An image generator seems to conjure pictures from thin air, but the process is mechanical and, once seen, demystifying. The dominant method is called diffusion. During training, the system takes millions of real photographs paired with text descriptions and gradually corrupts each image with random visual noise, step by step, until nothing but static remains. The model's entire job is to learn the reverse: given a noisy image and a description, predict a slightly cleaner version. Repeat that denoising skill many times and it becomes generation — start from pure random static, tell the model the description of what should be there, and let it remove noise step by step until a coherent picture emerges. The text prompt steers every step because the model learned, from those millions of caption-image pairs, which visual patterns go with which words. Work through an example: prompt a generator for a market stall in Mbare with baskets of tomatoes at sunset. The model has never seen your exact scene, but it has learned the visual statistics of market stalls, baskets, tomatoes, and sunset lighting from its training data, and the denoising process assembles those statistics into a new image that satisfies all of them at once. This explains the characteristic failures. The model composes typical patterns rather than consulting reality, so it produces hands with six fingers, text on signs that dissolves into letter-like shapes, and details that are plausible in general but wrong in particular — a street scene with architecture that fits no actual city. It also explains a serious fairness issue: places, faces, and cultures that are rare in the training photographs are rendered less accurately, so images of African settings sometimes drift toward stereotype because the training data was thin or skewed. The misconception to correct is that a generated image is a kind of photograph — a record of something. It is a statistical composite, closer to a very skilled illustration produced from memory of other pictures. That is why honest labelling matters so much: the image can be useful as decoration or illustration, but it is never evidence that anything happened, and presenting it as a photo of a real event is misinformation regardless of how harmless the intent.

A class group wants to post an AI-generated flood image beside a local safety message. A responsible response is: Label it as an illustration and pair it with verified local safety information. Correct. Clear labelling prevents people from mistaking an illustration for a current photograph. Use this case to separate what the technology contributes from what people contribute. The team should compare the intended outcome with a baseline where applicable, record important assumptions, and keep a clear route to correct or stop the process.

Ask whether the examples, data, and assumptions fit your school, company, or community. Generic demos often miss local names, laws, connectivity, and languages.

AI can create convincing pictures. Treat them as illustrations unless a trusted source confirms they are evidence of a real event.