ChatGPT or Gemini · Practical walkthrough

Plan an image sequence that keeps one subject consistent

Three attractive images for a campaign can still show three slightly different bicycles. Decide what stays fixed before you ask AI to change the scene.

Task by Task7 min read + practice

What you'll make

a short frame plan that separates what must stay the same from what each frame changes, then a sequence you check one frame at a time.

What you'll need

one clear reference image of the subject (a product photo or a fictional character you are allowed to use), the purpose of the sequence, and ChatGPT or Gemini with image editing. Allow 30–45 minutes for three frames.

Example result.

Frame	Scene change	Must stay the same
1	Bicycle outside the shop front	Teal frame, two wheels, lime front carrier, no rear box
2	Collecting an order at a café	Same
3	Delivering at a front door	Same

1. List what must stay the same

Write two lists. The first is the subject’s defining features: shape, colour, accessories and any visible details such as a handle or logo position. The second is what each frame is allowed to change: background, placement, light or a deliberate new viewpoint.

A filename or a repeated description doesn’t guarantee the same subject. If exact product pixels matter, such as packaging for a retailer, place the real product photo in a design tool instead of asking AI to redraw it each time.

Plan a [number]-frame image sequence for [purpose]. The attached image is
the original reference for the subject. List the features that must stay
the same in every frame. For each frame, give the one scene change and the
composition. Keep any text out of the images. Then write the request for
frame 1 and the checks I should make before frame 2.

[number] is how many frames you need. [purpose] is what the sequence tells the viewer, such as “how our delivery service works”.

Why this works: writing the fixed features down gives you a checklist for every frame, and asking for one scene change per frame stops the model changing several things at once.

What you should see: a table like the example result, plus a short request for frame 1.

2. Generate one frame at a time from the original reference

In ChatGPT, attach the reference image with the + button and paste the frame request. In the Gemini app, upload the image and describe the edit; Gemini can take more than one reference image, so you can add the accepted frame 1 as a second image, labelled as such. Each tool still drifts, so check every frame.

Attach the original reference for every frame. Don’t let the latest generated frame quietly become the reference: if frame 2 gains a feature, frame 3 will copy it faithfully.

Use the attached original reference for the subject. Frame [number] changes
only the scene: [scene change]. Keep [features that must stay the same]
exactly as in the reference. No new accessory, label or viewpoint.
Return one image.

[scene change] is the single change from your plan. [features that must stay the same] is the first list from step 1.

Why this works: one image per request gives you a natural checkpoint, and naming the reference every time stops errors compounding.

What you should see: one image with the new scene and a subject that matches your list. Compare it with the reference before you ask for the next frame.

3. Check the subject first, then the story

Lay the frames side by side with the reference. Check the subject against your list first, then check that each frame tells a different part of the story. A matching colour doesn’t make up for a different product shape, and perfect continuity doesn’t make up for three frames that say the same thing.

Why this works: checking the subject and the message separately stops a strong scene distracting you from a changed product.

What you should see: the same subject in three distinct scenes, in the order of your plan.

Check the result

Every defining feature on your list appears in every frame, unchanged.
Nothing has been added: no extra accessory, label, pocket or viewpoint the reference doesn’t show.
Each frame changes only its planned scene.
The three frames tell three different parts of the story.

The usual failure: a frame adds a plausible accessory, such as a rear box on a delivery bicycle. It looks right for the scene, so it’s easy to accept, and the next frame copies it.

Frame [number] added [feature] that is not in the original reference.
Go back to the original reference image and remove only that feature.
Keep this frame's scene unchanged. Return one image.

If the same fault keeps returning, reduce the change, for example by keeping the same camera angle, or place the real product photo into each scene in a design tool.

Worked example: a delivery-bicycle sequence

Fictional example. A local bakery wants three images for a page about its bicycle delivery service.

Reference: photo of the bakery's teal delivery bicycle, side view
Fixed features: teal frame, two wheels, lime front carrier, no rear box
Frame 1: outside the shop front, morning
Frame 2: collecting an order at a café
Frame 3: delivering at a front door

Frame 1 matched the list. Frame 2 came back with a lime rear cargo box as well as the front carrier: it looked natural for a collection, and nobody had asked for it.

The repair prompt went in as “Frame 2 added a rear cargo box that is not in the original reference”, with the original photo attached again. The second attempt kept the café scene and removed the box. Frame 3 was then generated from the original reference, not from frame 2.

The fault was believable. A rear box suits a delivery bike, which is exactly why the checklist caught it and a quick glance didn’t.
The original reference did the repair. Re-attaching it, rather than editing frame 2 again, stopped the box drifting into frame 3.
The scenes still differ. Shop front, café and doorstep each say something new, so the sequence earns three frames.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Make intentional changes explicit

If the story needs a new accessory, outfit or product variant, write it into the frame plan as a planned change, for example “Frame 3: the rider adds a rain cover to the carrier”. Then check that frame against the plan, not the original list, and make sure the change doesn’t carry backwards into earlier frames you regenerate later.

Keep a frame register for long sequences

For more than three or four frames, keep a simple table: frame number, request used, reference attached, result file and whether it passed the checklist. A contact sheet of all frames makes small drift easier to spot. The annotated comparisons guide shows how to mark differences.

Storyboard before you generate

Write the message of each frame in one line before generating anything. Two frames with the same message is one frame too many, and it is cheaper to cut it from a list than after an hour of generation. The explainer storyboard guide covers this in more detail.