Skip to main content
AI Image Creation

LESSON 4 OF 5

What they are bad at

BY THE END OF THIS LESSON

Recognise the failures worth designing around rather than fighting.

Some things are hard for structural reasons

4 of 4 to check carefully.
  • Check this: Words rendered inside the picture
  • Check this: Hands and things with a fixed count
  • Check this: Exact quantities
  • Check this: The same character across images

A few failures recur across image tools and across releases: words rendered inside the picture, hands and other things with a fixed count, exact quantities — five items when you asked for five — and keeping the same character or product identical across separate images.

These are not a prompting skill issue. Better wording helps at the margin; it does not make them reliable.

Design around them instead

Asking for words inside the picture against Set the text over it afterwards, drawn at their real relative sizes.

Asking for words inside the picture

Ask a poster with the headline on it

Set the text over it afterwards

Ask the image with no words, top third emptyThen editable, correctly kerned

The reliable move is to stop asking for the hard thing. Generate the image without any words and set the text over it afterwards — which you wanted anyway, because then it is editable and correctly kerned.

Crop out the hands. Ask for "several" rather than a number when the number is not load-bearing. Where identity really matters across images, plan for the fact that separate generations will differ.

Check before you ship, not after

3 stages, each leading to the next.
  1. Count the fingers
  2. Read any lettering
  3. Compare against a real photograph

Look at the specific things that go wrong: count the fingers, read any lettering character by character, count the objects you specified, and compare a product against a real photograph of it.

These errors are invisible at a glance and obvious once printed, and the difference between catching one and not is thirty seconds of deliberate looking.

WORKED EXAMPLE

FIGHTING A KNOWN WEAKNESS

A poster with the headline "Scheduling that works" in bold at the top, hands holding a tablet.

DESIGNING AROUND IT

A tablet on a warehouse desk, seen from above, no hands in frame. Leave the top third empty and plain — no text of any kind in the image. [headline set in the design tool afterwards]

Text inside a generated image and hands are the two most reliable failure modes. Removing both from the request and adding the headline afterwards gives an image that works first time and a headline that is editable and correctly spelled.

YOUR TURN

Rewrite a brief to avoid the known failures.

RUN THIS

Here is an image brief: [describe what you need]. Tell me which parts of it are asking for something image tools are structurally bad at — words in the image, hands, exact counts, consistency across images. Then rewrite the brief to avoid each one, and tell me what I should do outside the tool instead.

What a good result looks like

Anything involving text should come back as "generate without it and set the type afterwards". If the brief needs the same character twice, expect an honest warning rather than a promise.

KNOWLEDGE CHECK

Answer all 3 correctly to complete this lesson.

  1. 1. Why not solve text-in-image failures with better prompting?

  2. 2. What is the reliable approach to a known weakness?

  3. 3. Why check the specific failure categories deliberately?

REMEMBER THIS

Design around text, hands, counts and cross-image consistency rather than prompting harder.