How to Use Reference Images for Consistent AI Generations

A reference image gives an AI image workflow visual information that would be difficult to describe completely in words. It can establish a person’s identity, a product’s shape, a room layout, a pose, a palette, or the overall structure of a composition. The written prompt still matters: it tells the model which parts of the reference should remain stable and which parts should change.

Useful reference-image prompts divide the task into preserve, change, and finish. Preserve the elements that define success, change the specific elements required by the new concept, and describe the visual finish you want in the result.

This guide shows how to make that division clear and how to review the output for consistency.

When a reference image is the right tool

Use a reference when the source contains information that is both important and hard to reconstruct from text.

Identity and likeness

A portrait reference can help carry recognizable facial structure, hair, and other identity cues into a new setting or art direction. The prompt should focus on role, expression, wardrobe, lighting, crop, and background rather than redescribing the person from scratch.

Reference input does not remove the need to review likeness. Compare the generated result carefully, especially when the image represents a real person in a professional or public context.

Product shape and material

Product references are useful when silhouette, packaging proportions, controls, labels, or material transitions matter. A written prompt can then direct the scene: studio or lifestyle context, camera angle, lighting, supporting objects, and campaign mood.

Generated details and text still require inspection. If a logo, ingredient label, price, measurement, or regulatory statement must be exact, add or verify it in the final production file.

Composition and layout

A reference can establish pose, viewpoint, object arrangement, or the relationship between foreground and background. This is useful for recreating a layout with a different subject or exploring a new style without losing the original hierarchy.

Visual system and continuity

When producing a family of campaign images, a selected result can become the visual reference for the next variation. Reusing palette, lighting direction, materials, lens feel, and shape language helps new images belong to the same system.

If none of these constraints matter, text alone may be faster. A reference should solve a specific control problem rather than being added by default.

The preserve-change-finish framework

Reference-image prompting becomes clearer when the instruction separates three jobs.

1. Preserve

Name the elements from the source that must remain recognizable. Be selective and concrete.

Examples include:

• preserve the person’s identity, face shape, and hairstyle
• preserve the bottle silhouette, cap design, and matte steel material
• preserve the seated pose and three-quarter camera angle
• preserve the room layout and the window position
• preserve the palette and broad directional light

Avoid saying “keep everything the same” when you also want major changes. That creates an unclear boundary. List the characteristics that define continuity.

2. Change

Describe only the transformation you need. Common changes include background, wardrobe, expression, time of day, season, camera distance, or art direction.

For example:

Preserve the person’s identity, face shape, and short dark hair. Change the casual snapshot into a chest-up professional editorial portrait with a calm direct gaze, navy overshirt, warm gray studio background, and soft directional light.

Or for a product:

Preserve the uploaded bottle’s silhouette, cap, label proportions, and brushed steel material. Place it on a dark stone pedestal in a clean studio scene with cool rim light, subtle condensation, and a wide campaign composition with empty space on the left.

3. Finish

Finish describes how the generated image should feel and where it will be used. Include medium, texture, light behavior, crop, aspect ratio, and production needs.

Useful finish instructions include:

• realistic editorial photography with natural surface texture
• premium ecommerce lighting with controlled reflections
• tactile cut-paper illustration with restrained shadows
• vertical poster composition with clear type-safe zones
• square profile crop with an uncluttered background

Separating finish from the transformation keeps the prompt readable and easier to revise.

Prepare the source image

The reference itself affects how clearly the model can interpret the task. Choose a source that makes the preserved elements visible.

For portraits, use an image with adequate resolution, clear facial features, and lighting that does not hide important structure. For products, avoid heavy perspective distortion, motion blur, or objects covering the silhouette. For composition references, make sure the hierarchy you want to preserve is obvious.

Only upload images you may use. Be careful with private portraits, client assets, copyrighted campaign work, and branded materials.

Write the prompt around the reference

Begin by identifying the reference’s role. A concise structure is:

Use the uploaded image as the reference for [identity, product, composition, or visual style].
Preserve [specific invariant features].
Change [specific transformation].
Finish as [medium, light, palette, crop, and intended use].
Do not change [critical exclusions].

The exclusion line is optional. Use it for a small number of high-impact boundaries, such as “do not add extra packaging,” rather than a long negative list.

The AI image generator with reference image workflow provides a focused entry point for trying this structure.

Reference prompt examples

Professional portrait from a casual photo

Use the uploaded portrait as the identity reference. Preserve facial structure, eye color, hairstyle, and natural proportions. Change the scene into a chest-up founder portrait with a relaxed direct gaze, simple charcoal overshirt, and warm gray studio backdrop. Use soft directional window-style light, natural skin texture, restrained editorial retouching, and a square professional profile crop. Do not add accessories or change age.

Product campaign variation

Use the uploaded product image as the shape and material reference. Preserve the bottle silhouette, cap geometry, label placement, and translucent amber glass. Move it into a quiet bathroom campaign scene on pale limestone with a folded white towel and one small botanical accent. Use soft morning side light, realistic contact shadows, controlled reflections, and a wide ecommerce composition with copy space on the right. Do not alter the package proportions.

New style with the same composition

Use the uploaded image as the composition reference. Preserve the low viewpoint, central cyclist, diagonal road, and distant mountain scale. Reinterpret the scene as a layered screen-print illustration with bold flat shapes, deep blue coral and pale cream inks, subtle paper grain, and a vertical travel-poster format. Keep the road and cyclist as the primary visual path.

Campaign continuity from an earlier result

Use the uploaded campaign image as the visual-system reference. Preserve the cobalt and warm cream palette, hard upper-left key light, rounded modular shapes, and clean product-stage geometry. Create a new scene for the companion product with a different pedestal arrangement while keeping the same lens feel, material finish, and amount of negative space.

Evaluate consistency in the result

Review the generated image against the preserve list before judging style. A beautiful transformation is not successful if it loses the feature that made the reference necessary.

For identity

Compare face shape, relative feature placement, hairline, and other recognizable characteristics. Also check that the new lighting or expression has not introduced an unintended change in age or identity.

For products

Check silhouette, proportions, openings, controls, edges, material transitions, label position, and any feature that distinguishes the product. Review reflections and contact shadows to make sure the object belongs in the new scene.

For composition

Compare viewpoint, focal point, scale relationships, negative space, and the path the eye follows. Minor object differences may be acceptable if the hierarchy remains intact.

For a visual system

Place the new image beside the reference at the size where both will be used. Check palette, contrast, light direction, texture, camera distance, and spacing. Continuity is easier to judge as a group than one image at a time.

Common reference-image problems

The result copies details that should have changed

The preserve instruction may be too broad. Replace “keep the image the same” with a short list of true invariants, then state the transformation immediately after it.

The result loses identity or product shape

Make the reference role explicit and move the invariant features to the beginning. Reduce competing style instructions, and use a source where those features are clear.

The new background does not match the subject

Direct light direction, intensity, color temperature, contact shadows, and reflections. A background replacement feels convincing when the subject and environment share the same physical light.

Multiple references compete

Assign one role to each reference: identity, wardrobe, composition, material, or style. Do not assume the model knows which source should control which property.

Generated labels or text drift

Treat text as a production step. Use the generated image for the visual direction, then restore verified logos, labels, prices, and required copy in the final design file.

Build consistency through controlled reuse

Reference images work best inside a deliberate sequence. Start with a clear source, write the preserve-change-finish prompt, and review the invariant features first. Once a result establishes the desired visual language, use that approved result as the reference for the next variation.

Keep a record of the prompt, source role, canvas, and the decisions that produced the approved image. That record turns a single result into a repeatable creative system rather than an isolated experiment. When text direction alone is enough, return to the text-to-image workflow; when continuity matters, use the reference to make the boundary between stable and changeable elements explicit.