Keeping one character consistent across a set of images
Describing a face ten times gives you ten faces. Anchoring to one start image gives you the same one, ten times.
Cancel anytime before day 7 · no lock-in
Start with one image, not ten prompts
The mistake that breaks a set is generating each image from scratch. Even a careful prompt like a young barista with short curly hair and a green apron lands on a slightly different person every time, because the words leave room and the model fills it differently on each run.
The fix is to pick one image and make it the anchor. Generate a single strong portrait or product shot you are happy with, then build every other image in the set from that one instead of from the description alone.
This one image carries the details words cannot pin down: the exact face, the precise shade of a label, the shape of a logo. Get it right once and you stop re-rolling the dice.
Send the anchor as a reference, then change the scene
Once you have your anchor, pass it as a reference image on every new generation and describe only what should change. Keep the character or product fixed in the reference and move the world around it: same character, now sitting at an outdoor cafe table at golden hour.
With Figgshield the reference image is used in memory for that generation and never stored, and the files you generate sit on a private link for seven days before they are deleted. So you can run a whole set through your agent without leaving copies behind.
If you already generated the anchor with Figgshield, you can point new generations at it by its id instead of re-uploading, which keeps a long series tied to the exact same source.
Change one thing per image
Consistency comes from discipline in the prompt, not just the reference. Hold the description of the character steady word for word across the set, and only edit the part that should differ: the pose, the background, or the light.
If you rewrite the whole prompt each time, small wording shifts pull the result away from the anchor even with a reference attached. Treat the character description as a fixed block you paste unchanged.
A simple pattern: [same character block] + [new setting] + [new action]. Only the last two brackets move from image to image.
A worked example: a mascot across a product launch
Say you are launching an app and want a friendly fox mascot across a landing page, three feature cards, and a set of social posts. Generate the fox once, front on, clean background, until the character feels right.
Then produce the set from that anchor: same fox, waving, holding a phone for the hero, same fox, pointing at a chart for a feature card, same fox, celebrating with confetti for the launch post. The fox stays the same animal in every frame while the action changes.
Because you are working from one source, the whole set reads as one character instead of a family of near-lookalikes, which is what viewers notice first.
Re-anchor when the look starts to drift
Over a long series the results can slowly wander, especially once you stack several changes at once. When an image comes back a little off, do not fight it with more adjectives. Go back to the anchor and start the next generation from it again.
If the anchor itself is the weak link, regenerate a better one and rebuild from there. A few extra credits spent on a stronger source image saves you from re-rolling twenty downstream images.
Keep the id or the file of your best anchor somewhere handy. It is the single most valuable asset in a consistent set, more than any individual output.
Draft cheap, finish sharp
Explore at 1K or 2K, then render the keepers at 4K. On Figgshield a Nano Banana 2 image is 8 credits at 1K or 2K and 16 credits at 4K, so testing poses and scenes at the smaller size costs the same eight credits and leaves room to try more.
Figgshield bills at cost with no markup on the model, so building a set of a dozen consistent images stays cheap. The $5 7-day Pro trial includes 150 credits, enough to anchor a character and produce a full set around it.
When a still is right, you can turn it into a five-second clip by passing its id, so your anchored character can move in a short video too. A clip is 35 credits and one credit balance covers both stills and video.