You will be able to explain how image generators and image-reading models work, and what each one is bad at.
Before you share a picture made by an AI tool, read every word in it letter by letter, count anything that was meant to come in a set number, and look closely at hands and small details. When an AI tool reads a picture and gives you figures, check each figure against the original before you use it. Both checks follow from how the two kinds of tool work and where each one tends to slip.
Joanne runs a small bakery in Tampines (she and her week are an invented example). On Monday she asks an image generator for a poster for her Hari Raya cookie promotion, and within seconds she gets a warm, glossy photo of cookie jars on a festive table. It looks professional until she notices that the banner in the picture reads "Hari Raua Speical" and that one jar's lid is floating beside it. On Wednesday she photographs a crumpled supplier invoice and asks an assistant to add up the flour orders. The total she gets is neat and a little over what she expected. Later she finds that one quantity had been read as 18 instead of 10.
The poster and the invoice involve tools with opposite jobs. An image generator makes pictures from words. An image reader takes a picture and answers in words.
Many image generators use a method called diffusion. To train one, builders take millions of captioned images and add random noise to each, step by step, until nothing of the original remains, much like static on an old television screen. The model then learns the reverse task. Given a noisy image and its caption, it predicts which noise to remove to bring the image one step closer to a clean picture.
Lesson 2.2 described training as a loop of guess, measure the miss, adjust, and diffusion training follows the same loop. The model's guess is "this is the noise", and the miss is measured against the noise that was actually added.
When you ask for a new image, the generator starts from pure random noise. It removes a little noise, guided by your prompt, and repeats this over many steps until a picture forms. At each step the prompt pushes the image towards something that fits your words, so "cookie jars on a festive table" produces jars and a table rather than a beach. The generator does not copy or edit a stored photo. It builds a new image that matches patterns learned from its training images.
Not every generator uses diffusion, and builders often combine methods. In all of them the image is assembled from learned patterns, and at no stage does the model work from a plan of what should be where.
Lesson 3.1 showed that words are cut into tokens and turned into numbers. An image reader does something similar with a picture: it cuts the picture into small patches and turns each patch into a list of numbers, much like a word token. These image tokens go into the context window next to your question, and the model predicts a text answer in the usual way. It is still predicting text, and the text in front of it now includes your picture in a form it can use. For this reason you can upload a photo of a whiteboard, a chart from a report or a screenshot of an error message and ask about it in plain words.
Generators handle overall look well, including lighting, style and mood. They are weaker at anything that must be exactly right. Text inside an image is the classic weakness, because the model learned what banner writing tends to look like more than how words are spelled, which is how Joanne's banner came out as "Hari Raua Speical". Counts are unreliable too: ask for seven cupcakes and you may get six or nine. Fingers and joints have long been a weak spot. Consistency also slips, so if you ask for the same mascot in five different poses, its face, colours and outfit tend to drift between pictures.
These tools keep improving, and some of these faults are much less common than they were. Check every generated image for text, counts, hands and small details, such as a floating jar lid, before anyone outside your team sees it.
Image readers are handy for getting text and figures out of photos and screenshots. Their main risk is a confident misreading. Small print, crumpled receipts, blurred photos and handwriting all cause trouble, and dense tables cause the most. A 6 can become an 8, a column can shift by one row, or a smudged figure can be replaced with one that looks plausible.
When a value is hard to read, the model does not leave a gap. It predicts the likeliest text, and that may be a number that was never on the page. On Joanne's invoice, a 10 came back as 18, and her total was a little high as a result.
Treat any figure an image reader extracts as a draft and compare it with the original. Take most care with totals, dates, amounts in S$, and anything you will pay, claim or report. The checking routine from Lesson 6.3 applies to these figures in the same way it applies to text answers.
To try this, give an image reader something with lots of numbers, such as a receipt from your last grocery run or a chart from a work report. Ask it to list every figure it sees, then go through its answer line by line against the original.
Give an assistant a photo of a receipt or a chart and ask it to list every number it extracted. Then mark which ones it got wrong.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).