Generating one good picture of a character is easy. Generating a second one that is recognisably the same character is the hard part, and it is the wall most people hit as soon as they try to make a series — a comic, a set of illustrations, a storyboard, anything where a face has to recur.

The reason is that a prompt describes a type, not an individual. “A young woman with short dark hair” matches millions of faces, and the model picks a different one every time. Consistency comes from narrowing that space, and there are two ways to do it: with words, and with a reference image.

Write the character down and never change it

The word-based approach is to fix a description once and paste it into every prompt, byte for byte. Not paraphrased, not reordered, not trimmed for length — the same block of text each time. Small rewordings shift the result more than seems reasonable, so treat the description as a constant rather than something to keep improving.

What goes in it should be specific and stable. Age range, hair length and color and how it is worn, face shape, eye color, skin tone, a distinctive feature or two, and a fixed outfit. The distinctive feature does most of the work: a scar, heavy round glasses, a particular jacket, an unusual hair color. Ordinary faces average out; one memorable detail gives the model an anchor to keep hitting.

Keep everything else out of the block. Pose, expression, camera angle, setting and lighting all belong in the part of the prompt that changes. When those get mixed into the character description, you end up locking the wrong things — the same pose every time, and a face that still drifts.

Reference images do it better

Where the tool supports it, feeding in a previous image of the character is far more reliable than any description. The workflow is straightforward: generate until you get one image you are happy with, then treat that image as the source of truth and reference it for every subsequent one, rather than referencing the most recent output.

That last part matters. Chaining each new image from the previous one lets small changes compound, and after a dozen steps the character has quietly become someone else. Always going back to the same original keeps the drift bounded.

A useful trick when starting a series is to generate a character sheet first — the same person shown from several angles and with a couple of expressions, in a single image. It gives you a reference that covers more than one view, which makes profiles and three-quarter angles much steadier later on.

Accept where it breaks

Some things will not hold. Fine facial detail changes across large shifts in angle and distance; a close-up and a wide shot of the same character will rarely match perfectly. Clothing details multiply and mutate — the fewer items and patterns in the outfit, the better it survives. Hands and jewellery are unreliable everywhere and no more reliable here.

The workable standard is not identity but recognisability. If a reader can follow the same character through a sequence without stopping to wonder, the series works. Chasing a pixel-level match will burn far more attempts than it is worth.