Keep a character or product consistent across scenes
Pin down what must never change in a character sheet, then use a fixed block and reference images to keep the same look in every scene.
14 minAs of Oct 2026
By the end of this guide
- Build a sheet that separates what stays fixed from what changes, for a character or a product.
- Condense the sheet into a fixed block and use it with reference images to keep the same look across scenes.
- Compare results with the sheet, fix only what's wrong, and finish products, logos and text in an image editor.
Who this is for
- Designers who need a character or mascot in a series of scenes
- Marketers who need the same product in many different settings
What you need
- Knowing the fixed-block method from "Make a mood board in ten minutes" helps.
- An image generator that accepts reference images
- A chat assistant that can read images, and an image editor such as Photoshop
Step by step
Make a sheet of what's fixed and what changes
Put the same character into several scenes and the eyes, outfit colors and marks drift a little from image to image, because each generation re-imagines whatever the prompt doesn't spell out.
So start with a character sheet. Separate what must never change — shape, proportions, colors, outfit, distinctive marks — from what changes every time: scene, pose, expression, light. Give colors as hex codes, and say whose left and right you mean, so you have something to compare against later.
Character sheet for Puff, the mascot of a fictional weather app (example) Type Item Details Fixed Shape Round, cloud-shaped body, two short legs, two small half-circle arms Fixed Proportions Width to height (legs included) 5:4; legs one-fifth of the height Fixed Colors Body #F8FAFC, cheeks #FDA4AF, raincoat #FACC15, boots #1E3A8A Fixed Outfit Hoodless yellow raincoat, navy boots Fixed Marks Two tall black oval eyes; a navy teardrop mark under the cheek on the viewer's left Changes Scene and props Set each time (e.g. a bus stop, an umbrella) Changes Pose and expression Set each time Changes Light and time of day Set each time Check it
- Could a colleague draw the same character from the fixed rows alone? Replace anything open to interpretation, like "cute eyes," with a shape and a position.
Condense the sheet into one fixed block
Write every prompt as a fixed block plus a changing block. The fixed block is the sheet's fixed rows condensed into one paragraph, pasted in unchanged every time. The changing block holds only the scene, the pose and expression, and the light, as in the table below.
A chat assistant can condense the sheet for you. Give it a length limit, and the fixed block stays short enough to keep separate from the changing part.
Changing blocks that follow the fixed block (example) Scene Pose and expression Light A bus stop in the rain Holding an umbrella, on tiptoe, smiling Soft light on an overcast afternoon A windy hill Gripping a kite string with both arms, surprised Bright sunshine on a clear morning A snowy alley Looking up at the snowflakes, mouth open Warm glow of streetlights in the evening Example prompt Condense only the 'Fixed' rows of the character sheet below into one paragraph I can paste into every image prompt. Keep it to 50 words or fewer, keep the hex codes as they are, and leave out the 'Changes' rows. [Paste the character sheet]
Example AI reply Round, cloud-shaped mascot; width to height 5:4; two short legs (one-fifth of height) and two small half-circle arms. Body #F8FAFC, cheeks #FDA4AF, hoodless yellow raincoat #FACC15, navy boots #1E3A8A. Two tall black oval eyes; navy teardrop mark under the cheek on the viewer's left.
Check it
- Count the length. The example comes to 44 words, within the limit of 50.
- Make sure all four hex codes and the position of the teardrop mark are there, and that nothing from the 'Changes' rows — scene, pose, expression, light — slipped in.
Make reference images first, then attach them
Use the fixed block to make one full-body front view and one three-quarter view (turned 45 degrees), each on a plain background. Generate several and pick the ones that match the sheet best as your reference images.
From then on, attach the reference images whenever you make a scene. Many tools accept reference images, but the feature's name, how many images it takes and what it copies (a character, an object, a style) vary by service and model. The Gemini API docs, for example, say that depending on the model it accepts up to four or five images for character consistency, and that including previously generated images in the next prompt helps keep a character consistent (as of October 2026).
Example prompt [Paste the fixed block] Full body, standing straight and facing forward, plain light grey background, soft even lighting
Check it
- Check the reference images against the sheet, item by item. Anything wrong here spreads to every later scene, so regenerate until they match.
- Keep the reference images in one place, with the date and a version number in each file name.
Compare with the sheet and fix only what's off
When a scene comes back, compare it with the sheet. A chat assistant that can read images will spot differences quickly if you give it the image and the sheet together — but it can miss things or misread them, so do the final check yourself.
If only one or two things are off, fix just those instead of regenerating the whole image, which can change the parts that were right. Ask in the conversation to change only that part and keep everything else, or use an editor feature such as generative fill, where you paint over an area and regenerate only that.
Example prompt List where the attached image differs from the character sheet below, as a table. Use the columns item, sheet, image and fix, in that order, and leave out anything that matches the sheet. [Paste the character sheet]
Example AI reply I found three differences.
For something that's only in the wrong place, like the mark, try: "In this image, remove only the teardrop mark and draw the same navy teardrop mark under the cheek on the viewer's left. Keep everything else exactly the same."
Differences the assistant found (example) Item Sheet Image Fix Raincoat color Yellow #FACC15 Yellow leaning orange Correct only the raincoat color in an editor Teardrop mark Under the cheek on the viewer's left Under the cheek on the viewer's right Ask to remove it and redraw it on the left Arms Two small half-circles Arms with fingers Paint over the arms and regenerate only them Check it
- Look again at each difference yourself. For colors, a value sampled with your editor's eyedropper beats the assistant's estimate.
- After a partial edit, zoom in around the fix to check for awkward edges or anything else that changed.
Start products from a real photo; add logos and text by hand
Products are stricter than characters: a slightly different bottle shape or label makes it a different product. Rather than generating the product, give the model a real product photo and ask it to change only the background, props and light. The more precisely you describe what must stay (shape, label, colors), the more likely it is to survive.
Some models render logos and text fairly well, but matching the exact typeface and shape is hard. Add the original logo file and the text in your editor at the end, and match the colors to your brand values there too.
Who does what in a product scene Element How Why Product shape and label Use the real product photo as is Generated details tend to drift Background, props, light Give the photo and describe only what should change The part a generator handles well Logo and text Add the original files in your editor Generators may not match the typeface and shape exactly Final colors Match brand color values in your editor Colors in generated results tend to shift slightly Example prompt Keep the tumbler in the attached photo exactly as it is — don't change its shape, label or colors. Change only the background to a café window on a rainy day, and make the shadows match the soft, overcast light coming through the window.
Check it
- Put the product in the final image next to the original photo and zoom in to compare the shape, label and lid color.
- Overlay the logo and text on the original files to check that the shapes match.
Common mistakes
- Rewriting or tweaking the character description for every scene
- Regenerating the whole image to fix one detail, and losing the parts that were right
- Putting a logo or text drawn by the generator straight into the final artwork
References
- Nano Banana image generation — Google AI for Developers
- Use generative fill — Adobe Help Center
- Reference images for styling — Adobe Help Center