Learning AI image generation is mostly a matter of unlearning a few wrong assumptions about how the tools work.
The eight mistakes below look unrelated but share a root: a wrong assumption about the mechanism. Correct the assumption and the mistake disappears on its own.
Each one, with its cause and its fix.
1. Cropping to the aspect ratio afterwards
Generating a square and cropping it wide risks cutting away the most important part of the composition. The model builds the scene around the ratio you give: 16:9 reads a landscape as wide, 9:16 reads the same landscape as a vertical slice.
The fix is simple: state the ratio in the prompt or the interface. Instagram post 4:5, story 9:16, YouTube thumbnail 16:9, poster 2:3.
2. Trying to render all the text inside the image
A concert flyer carries the band name, date, venue, time and ticket price. No model writes all of that without errors; past three words the error rate climbs sharply.
Split the job: render only a one- or two-word headline in the image, leave empty space in the composition for the rest, and add the copy afterwards in a design tool. That gives error-free typography and means a date change does not force a regeneration.
3. Not checking resolution for print
A poster that looks perfect on screen can print blurry, and the cause is pixel count rather than design. Print wants 300 DPI: an A3 poster needs roughly 3500×4900 pixels.
Most tools generate between 1024 and 2048 pixels — plenty for screen, not enough for A3. Run the output through an upscaler and inspect at one hundred per cent zoom before printing, paying particular attention to the edges of letters.
4. Generating your own product from scratch
The jar the model generates resembles your jar, but the label differs, the lid has different proportions, and your brand is not written on it.
In e-commerce the image shown must accurately represent the product sold; showing something different raises returns and creates regulatory exposure.
The correct approach is the reverse: you shoot the product, AI generates the world around it. Remove the background, generate a new scene, place the product into it — and make sure the light directions agree, because that is the single detail that determines believability.
5. Naming living artists in prompts
Naming an artist is the known shortcut to their style, and that is exactly the problem. In commercial work it is ethically indefensible, it breaches some platforms' terms, and most tools already filter those names.
The practical argument is more persuasive: a name gives you that artist's average, not control. Describe the technique — line weight, colour count, texture, shading method — and you can tune the result.
6. Solving consistency by repeating prompts
The same prompt run twice gives two different images; that is richness for one image and a problem for a feed of ten.
Prompt discipline helps but is not sufficient. The real fix is style reference: Midjourney's --sref parameter, saved styles in Leonardo and Recraft. If a character recurs, you need --cref or a LoRA trained for that character on the Stable Diffusion side.
7. Uploading unoptimised files to the web
The files these tools hand you are usually PNGs over a megabyte. Three images in an article puts the page at three megabytes on imagery alone.
Page speed is a ranking factor, and on most article pages the Largest Contentful Paint metric is bound to the header image. Convert to WebP and resize to what you actually need; 1200 pixels wide is plenty for a blog.
8. Skipping the copyright check before commercial use
Because models were trained on existing imagery, the chance that your output resembles a recognisable character, brand or artwork is not zero. A t-shirt design or ad visual can be pulled from a platform for exactly this reason.
A two-minute precaution covers it: reverse image search the result before listing it. If something similar comes up, change the design.
For corporate work there is a further layer: using a model trained on licensed data. Adobe Firefly is in that category and offers indemnification on enterprise plans.
Frequently asked questions
Why does my image look fake?
Almost always because the light does not match. If you placed a product into a generated scene, the shadow direction in your product photo and the light direction in the scene must agree, and there must be a soft contact shadow where the product meets the surface.
Why are hands still broken?
Models do not know as a rule that a hand has five fingers; they learn statistical patterns. Flux is markedly better here. Choosing framing that keeps hands out of shot and adding "extra fingers, deformed hands" to the negative prompt solves most of the rest.
How many attempts before I give up?
If you are not converging after five or six, the problem may be the tool rather than the prompt. Designs containing text belong in Ideogram; work requiring a specific preserved pose belongs in Stable Diffusion with ControlNet.