Keep AI Image Iterations Consistent

A series is a different job from an image

Producing one good image and producing twelve that belong together are different problems, and the second one is mostly bookkeeping.

The reason is worth understanding. Each generation is an independent sample. Nothing carries between them unless you carry it. Two runs of the same prompt produce two images that a viewer will read as coming from different shoots, because the light direction moved, the palette drifted, and the camera height changed — none of which you specified, so the model chose each time.

So the work is to decide which variables are locked for the whole set and write them down once. The locked list is not a style preference; it is the definition of the set. Keep it as a literal block of text that gets pasted into every generation:

SET SPEC v3  "autumn kitchenware", 12 images
────────────────────────────────────────────
LOCKED   matte white ceramic, no glaze variation
         single soft key light from camera-left, 45 degrees
         warm neutral background, RAL 9001, no gradient
         camera at object height, slight downward tilt
         4:5, object occupies 60-70% of frame height
         no props, no hands, no text

FREE     object type, object rotation, shadow length
         within the range shown in reference set A

Two details make this work in practice. The spec is versioned, so when something changes mid-project you can tell which images came from which rules. And the free variables have stated ranges rather than being genuinely free — "shadow length within the range shown in reference set A" is a constraint, and without it the free column silently becomes a second locked column that nobody agreed on.

One variable per round

When a result is close, the temptation is to fix everything at once. Resist it, for a reason that is more practical than methodological: if you change three things and the image improves, you have learned nothing you can reuse.

Record the full state first — seed, prompt version, reference image, model version, and every sampler setting. This is the point you can return to, and you will need to return to it. The most common frustration in image work is having produced something excellent two hours ago and being unable to get back to it.

Then change one variable, compare side by side rather than from memory, and write one sentence about what the change did.

That sentence is the actual output of the round. The image is disposable; the knowledge is not. After a dozen rounds you have a page of sentences like "increasing style strength above 0.6 loses the matte finish" and "camera below object height makes the handle read as larger than it is," and the next project starts from that page.

Memory, incidentally, flatters the most recent image. Side-by-side comparison at the same size is not a formality — reviewers reliably prefer whichever image they saw last when they compare sequentially.

Select against the set

The final step is the one that most affects whether the work looks professional, and it is a selection problem rather than a generation problem.

View the candidates together and small. At thumbnail size, a break in palette, light direction, or crop is immediately obvious; at full size it is invisible, because you are looking at the subject rather than the pattern. Print them, or build a contact sheet — the constraint of seeing all of them at once is what does the work.

Then find the outlier. There is nearly always one image that is individually the strongest and does not belong: a slightly different light, a more dramatic angle, a moment of genuine luck. It has to go. Cutting it feels wrong and improves the set more reliably than any other single decision, because a viewer reads inconsistency as a mistake even when they cannot name what changed.

If the set will be seen in sequence rather than as a grid, check each adjacent pair as well. Two images that are individually fine can produce a visible jump between them.

Finish by updating the spec with what actually held. The spec you wrote at the start is a hypothesis; the spec you write at the end is documentation, and it is what lets someone else extend the set in three months without the result looking like a different project.

Practice: a six-image contact sheet

Pick one brief and produce six related images under a written spec.

Work in rounds of one variable. Keep the sentence log. When you have twelve to fifteen candidates, build a contact sheet, select six as a group, and deliberately cut your favourite individual image if it does not belong.

Then hand the final spec and one of the six to a colleague and ask them to produce a seventh that fits. Whether they can is the real test of the spec, and the gap they hit is the line you forgot to write.

Common mistake

Do not use reference images you do not have permission to use.

This is the mistake with the longest tail, because reference images influence output in ways that are hard to see afterwards. A reference can carry a style, a composition, a recognisable identity, or a protected character into the result even when the final image is not a literal copy and even when the resemblance is not obvious to you.

Track what went into the set the same way you track the spec: which references were used, where they came from, and what their terms allow. Reconstructing that six months later — when a client asks, or when the set is reused for a different campaign — is close to impossible, and "we are not sure what we referenced" is not an answer anyone accepts.