A useful reference-image prompt identifies the source subject, names the details that must remain recognizable, and describes the change separately. Add a short priority rule when those requirements could conflict. “Keep the subject exactly the same, but make everything more creative” leaves the most important decision unresolved.
- Start with the kind of change you actually want
- Build the prompt around five decisions
- What a real submitted prompt achieved—and what it did not
- Three worked examples of clearer keep/change wording
- Match the wording to the control you are using
- Resolve contradictions before making the prompt longer
- Revise from the visible failure
- Use the prompt as a brief and the image as the evidence
In Dreamina, upload the reference and write the change as a specific visual task. For example, ask for the central bottle on a gray studio surface, then name its open neck, green glass and red star as identifying features. That gives both generation and review a clearer target. It does not turn a generative rendering into guaranteed pixel preservation.
The aim is a prompt you can use, inspect and revise—not a long list of impressive adjectives.
Start with the kind of change you actually want
“Use this reference” can mean several things. You may want a new scene containing the same subject, a different rendering of the same composition, or a small correction to an existing image. Those tasks need different instructions and sometimes different controls.
Do not describe a whole-image transformation as if it were a tiny edit. Turning a photograph into a flat illustration necessarily changes surface texture and shading. Replacing a background may also require new contact shadows. Name those permitted changes instead of saying that every visible property must stay fixed.
If several reference images are involved, choose their responsibilities first. The subject, composition and style reference guide explains how to avoid asking a style example to redefine the subject or a subject photograph to override the intended layout.
Build the prompt around five decisions
These are writing responsibilities, not mandatory labels or a special platform syntax. A natural paragraph can express all five.
Identify the exact subject in the input
When the source has multiple objects, distinguish the intended one. “The complete central green bottle” is clearer than “the bottle” when a second bottle is partly visible. “The person on the left wearing the blue jacket” is more specific than “the person” in a group photograph.
Describe only what the reference actually shows. If the back of a jacket or the other side of an object is invisible, an instruction to preserve it cannot supply the missing information. Use another authorized view or accept that the model must infer it.
Name the few details that define success
Choose visible properties you can check: silhouette, an open rather than capped neck, a distinctive color area, a clothing feature or a specific pose. Those are more useful than “same quality” or “same feel.”
Distinguish identity from measurement. “Retain the long neck and broad lower body” describes recognizable form. “Do not change any proportions” expresses a stricter intention, but it does not establish that geometry will actually remain identical. If measured dimensions are required, choose a workflow that retains or controls them directly.
For lettering, decide whether it must be legible, exact or merely visually suggested. Those are different standards. A small illustration may use simplified label areas; an actual package image needs accurate copy. The prompt should not quietly substitute one for the other.
Describe the allowed change in observable terms
“Make it premium” could affect the subject, background, lighting, props and composition at once. “Place it on a pale stone surface with soft side light and an uncluttered background” defines a result you can inspect.
Similarly, “make it illustrated” can mean anything from a painterly photo to flat geometric artwork. Specify the treatment that matters: broad shapes, limited colors, paper texture or reduced shading. Keep unrelated style terms out of the first request.
State which requirement wins a conflict
If the subject must remain recognizable while the style simplifies detail, say which details are essential and which can be abstracted. If layout is required for a headline, keep that area clear even if the preferred style example contains dense decoration.
Priority wording does not reveal or directly set a model's internal attention weights. It makes your creative decision explicit. That matters when you review the result: you know which failure requires correction and which variation is acceptable.
Exclude plausible distractions
Name things that could reasonably enter from the input or brief. If the source includes tableware and a partial bottle, exclude them. If a portrait should remain a single-person image, exclude additional people. If the output is a clean visual for later layout, ask for no added words or signs.
Avoid a huge generic negative list. It can distract from the actual subject and make revision harder. A useful exclusion has a clear relationship to a likely error or the final deliverable.
What a real submitted prompt achieved—and what it did not
On October 14, 2026, a restaurant photograph containing a complete central bottle, a partial bottle and tableware was uploaded for this demonstration. The generation used Seedream 5.0 Pro on a Basic account, a 3:4 ratio, Standard 1.5K and one output, downloaded as a 1296 × 1728 JPEG. The analysis uses the saved prompt, settings and original output.
This is the exact submitted prompt:
Use only the complete central green glass bottle in the reference. Place it on a neutral light-gray studio surface. Keep the open uncapped neck, bottle proportions, three labels, visible brand lettering, red star and 500 ml marking. Exclude the partial bottle and all tableware. Soft even lighting, photorealistic, no extra text.
The output removes the original table setting and retains recognizable bottle features, while redrawing fine details. Source photograph: Philafrenzy, “San Pellegrino 500ml bottle,” Wikimedia Commons, CC BY-SA 4.0. This AI-modified adaptation is shared under the same license. Product branding identifies the example and does not imply endorsement.
The opening sentence identifies the subject. The second specifies the new environment. The “Keep” sentence lists the intended invariants, and the exclusion sentence resolves which surrounding content should disappear. The finish instructions describe the desired lighting and rendering.
The resulting bottle retains its green glass appearance, uncapped neck, three label areas, red star and readable 500 ml marking. However, small lettering, label curvature, perspective and highlights are redrawn. That distinction matters: the word “Keep” states the requested constraint; the image establishes what was actually retained.
This example supports a reference-led scene change with recognizable features. It does not establish that this wording works universally or that the original bottle pixels remained untouched. For exact product representation, preserve approved product artwork or use an appropriate compositing route.
Three worked examples of clearer keep/change wording
The following examples are newly written instructional prompts. They were not submitted in the demonstration above, and that result image does not validate them.
Example 1: change the scene without redesigning the object
Vague request: “Make this bag look luxurious in a beautiful setting. Keep it the same.”
The problem is that “luxurious” may encourage changes to the bag's material, hardware or silhouette. “The same” does not say which of those matter.
More specific brief:
Use the complete tan shoulder bag in the reference as the subject. Retain its curved flap, single shoulder strap, rectangular brass clasp and plain tan leather appearance. Place it on a pale stone ledge against a quiet warm-gray background. Change the setting and supporting shadows; do not add pockets, patterns, lettering or hardware. Prioritize the bag's existing design over decorative styling. Keep the full bag visible, with enough space around it for a square crop.
The useful change is scope. The new environment carries the premium mood, while the bag's visible design remains the identity target. The crop requirement also prevents the scene from becoming attractive but unusable for the intended placement.
If the real item must be literally unchanged, this remains a brief rather than a guarantee. Use the original product layer for the final composite instead of asking a new render to certify its own accuracy.
Example 2: change the medium while keeping the person recognizable
Vague request: “Turn this portrait into a minimalist illustration, keeping every detail exactly.”
Minimalism removes detail. Requiring every photographic feature to remain exact conflicts with the intended medium.
More specific brief:
Create a flat editorial illustration of the person in the reference. Retain the chin-length dark haircut, round glasses, red jacket and three-quarter pose as recognition anchors. Use broad shapes, an off-white background and a limited red, charcoal and cream palette. Simplify skin texture, fabric detail and shading; keep the glasses and jacket distinct. Do not add another person or text.
The prompt names what carries recognition and what simplification is allowed to remove. It does not require photographic pores and reflections inside a flat illustration. For a consented real-person portrait, inspect identity carefully rather than treating familiar clothing as proof that the face is accurate.
If you also need a matching photo-style version, use the photo-and-illustration workflow to choose whether both versions should start from the original or whether an approved composition should become the next reference.
Example 3: change the lighting without changing the room's arrangement
Vague request: “Keep this room identical but turn it into a dramatic evening scene.”
Evening light changes shadows, reflections and visible color. “Identical” may obscure which parts of the scene are actually fixed.
More specific brief:
Keep the room's visible arrangement: sofa against the rear wall, window on the left, low table centered in front of the sofa and the same clear path beside it. Change the scene to evening with a warm lamp glow from the right and subdued light at the window. Preserve the furniture designs and positions; allow shadows, reflections and apparent brightness to change with the new lighting. Do not add furniture, people or decorations.
This separates geometry from illumination. If the result moves the table, it has violated the arrangement. If a shadow changes, that may be necessary for the requested lighting. The prompt gives you a practical basis for distinguishing an error from the intended transformation.
Match the wording to the control you are using
Dreamina's reference-image workflow combines an uploaded image with written directions and output settings. Its Seedream 5.0 Pro guide also emphasizes subject, style, lighting and composition. Those categories are useful starting points, but the important addition for a reference task is identifying which category must stay and which should change.
Other applications distinguish generation, reference inspiration and editing explicitly. Midjourney's Image Prompts guidance asks for a description of the desired final image, while its Edit Model accepts change instructions. Translate the creative brief into the mode's intended language rather than assuming one paragraph is a universal command format.
Likewise, a selection tool supplies spatial information a sentence may not. It identifies the intended region, while your text describes the change. It still requires review: ChatGPT's image editor documentation notes that edits can extend beyond the highlighted area.
Use an actual aspect-ratio control when one is available. Asking for a square composition in text can express intent, but the output setting determines the selected canvas. Prompt wording and application settings should support the same task.
Resolve contradictions before making the prompt longer
Several common instructions contain conflicting requirements:
Adding “very important” to both sides does not resolve the conflict. Choose the real requirement and rewrite the other as a permitted change, a preference or a finishing step.
If an exact requirement cannot be achieved reliably through generation, change the operation. Typography can be added as editable text; product details can remain in original artwork; exact placement can be controlled on a canvas. A prompt should describe the task the tool can perform, rather than conceal a production requirement it cannot guarantee.
Revise from the visible failure
After generating, compare the result with the source and your stated priorities. Identify the first important failure before changing wording.
If the wrong subject was chosen, clarify its identity or simplify the source crop. If the subject changed, check whether the reference contains the necessary detail and whether the style request contradicts it. If unwanted objects appeared, name the relevant exclusion. If the composition is wrong, describe the specific relationship or use a spatial control.
Make the next revision address that problem. Do not simultaneously change the model, palette, viewpoint, background and prompt length unless you are intentionally starting a new direction. Otherwise, you will not know which change helped or which established detail was lost.
For a hypothetical bag result with a round clasp instead of the required rectangle, first check that the source shows the clasp clearly. If it does, identify the precise mismatch: “Restore the rectangular brass clasp shown in the source; retain the approved bag silhouette and stone setting.” That is a targeted revision example, not a tested correction. If the source hides the clasp, repeating the instruction more emphatically will not provide the missing visual information.
Then inspect the corrected version against both the source and the approved result. A rectangular clasp is not a successful repair if the strap or flap has changed in an unacceptable way. Keep the earlier approved file so you can distinguish the intended correction from new drift. If the small detail remains unreliable and is essential to accurate product representation, retain the original artwork or finish that detail manually. The stopping decision follows the requirement, rather than an arbitrary number of prompt rewrites.
For a series, keep the stable portion of the brief and change the frame's actual job. The coherent image-series guide explains how to vary action, scale and information while retaining the visual language. Reusing identical wording for every frame can create repetition; rewriting every fixed detail can create drift.
Use the prompt as a brief and the image as the evidence
Upload an authorized reference in Dreamina, identify the intended subject, state the necessary fixed details and describe the allowed transformation. Set the output format, generate, then inspect the result against those decisions before downloading the approved version.
Save the prompt, reference, settings and final file together. A later revision should begin with a clear explanation of what was approved, not a guess based on the image alone. Clear keep/change rules make that explanation possible—and make it easier to see when a different editing or finishing method is needed.