Edit One Part of an AI Image and Check What Stayed Unchanged

Dreamina's Inpaint workflow offers a practical way to make a selected-area change. In the recorded example below, one operation added a pumpkin to a gray studio image while keeping the bottle recognizable. The downloaded results also show why “recognizable” and “every unselected pixel stayed identical” are different claims.

This guide takes the edit through comparison, correction and final export, so you can approve the whole image with a clear reason.

Decide what “unchanged” needs to mean

An image can look similar at a glance while its lettering, outline or texture changes. Conversely, two files can differ numerically because of export processing while depicting the same visible content.

Choose the requirement before editing:

For a playful editorial image, small texture changes may be acceptable. For packaging, an altered letter can make the image unusable. For a scientific or documentary image, a creative reconstruction may be inappropriate even when it looks convincing.

Do not let a general instruction such as “keep everything else unchanged” stand in for this decision. It expresses the intention, but it does not establish what the tool actually retained.

Plan the edit around a target, a boundary and protected details

Name the target precisely: add a small prop in the empty upper-right background, remove a sign on the wall, correct a sleeve color or replace a particular object. Avoid combining an object change, new lighting and a style conversion when you only need a local revision.

Next identify the boundary. Adding a grounded object may require a contact shadow outside its silhouette. Removing an object may require rebuilding the surface behind it. Changing a sleeve can affect the transition into the hand. The necessary edit region is sometimes larger than the visible defect, but it should still have a reasoned scope.

Finally list the protected details that would make a good-looking edit fail. They might include a face, logo, product edge, headline, horizon or furniture arrangement. Keep the list short and specific enough to compare afterward.

For example, “replace the small background sign; keep the person's face, jacket and pose as approved” provides a clearer inspection task than “improve the background without changing anything.” The keep/change prompt guide explains how to express those priorities without turning them into conflicting instructions.

A recorded Dreamina edit: an addition inside a selected area

In the October 14, 2026 demonstration, Inpaint was applied to a brushed upper-right background area in a previously generated bottle image. One submission returned four images. Each downloaded JPEG was 1296 × 1728 pixels, matching the input dimensions. The analysis uses the saved prompt, settings and original outputs.

The source image had been generated with Seedream 5.0 Pro on a Basic account, at 3:4 and Standard 1.5K. The Inpaint submission dialog did not independently show the model behind that operation, so the inherited source label should not be read as a separate verification of the editing model.

The exact edit prompt was:

Add a small orange pumpkin only inside the selected background area. Keep the bottle, labels, lighting and all unselected areas unchanged.

Before and one of four returned images. The third output is shown to make the pumpkin's shadow and placement easy to inspect; it is not a claim of perfect integration. Source photograph: Philafrenzy, “San Pellegrino 500ml bottle,” Wikimedia Commons, CC BY-SA 4.0. These AI-modified adaptations are shared under the same license. Product branding identifies the example subject and does not imply endorsement.

All four outputs added an orange pumpkin in the intended upper-right area. Their placement differed: outputs one, two and four looked suspended against the background, while the third included a visible rightward shadow and a stronger suggestion of a surface. A shadow makes contact more legible, but the scene still needs a placement judgment; the request had not defined a tabletop or support at that height.

The bottle remained recognizable in all four, including its green body, uncapped neck, major label areas, red star and 500 ml marking. Fine lettering and edges were softer than in the input. That matters if the bottle's approved label detail is part of the deliverable.

The example therefore supports a selected-area addition with limits. It does not support a blanket statement that the rest of the image was preserved pixel for pixel, nor does one operation establish a general success rate.

What a numerical comparison added

The saved files were also compared as decoded RGB images at their original dimensions, without scaling or alignment. The comparison excluded a conservative upper-right rectangle larger than the visible brush selection. Outside that rectangle, about 4.39%–4.49% of pixels differed from the input by more than 20 levels in at least one color channel.

This is a file-difference measurement, not a percentage of the image that was semantically wrong. JPEG re-encoding and image processing can contribute to differences. The exact brush mask was not available as a separate file, so this is not a precise mask-leakage measurement either.

The useful conclusion is narrower: the exports were not numerically identical outside the broad target exclusion, and visual review identified softer detail worth checking. Neither a nonzero difference map nor a familiar-looking bottle alone answers whether the edit is acceptable for your particular use.

Make a focused edit in Dreamina

1. Keep an approved baseline

Download or preserve the image before editing. Give it a name that distinguishes it from later results. If you have already approved the face, label or composition, keep that version available throughout the process.

Avoid using a compressed screenshot as the baseline when an original export is available. Screenshots can change dimensions and color appearance, making a later comparison less meaningful. They are useful for recording the selected area or interface, but they serve a different purpose from the image file itself.

2. Choose the smallest complete operation

Open the available regional editing controls for the image. In the recorded Dreamina operation, Inpaint supplied brush selection and a written edit instruction. Select the area that needs reconstruction or addition, including necessary local context while avoiding important protected content where possible.

The selection should fit the task. An isolated decorative prop needs room for its silhouette and integration. A removal may need its shadow included. A color correction on an existing object may need a different route from inventing a replacement object.

Dreamina's photo-editing page documents prompt-led changes and refinement. Use the controls shown in your current workspace; the existence of a general editing feature is not proof that every model exposes the same selection tools or preserves unselected pixels in the same way.

3. Write the change and the protected content separately

Describe the requested result in visible terms. For a new prop, include its location, scale and relationship to the scene. For a correction, identify the defect and desired replacement. Keep the protected features in a separate sentence so the intention is easy to review.

A new instructional example could be: “Add one small ceramic cup resting on the empty right side of the table, with a contact shadow matching the existing light. Keep the person, table edge and approved text unchanged.” This example has not been tested. Its value is that it defines contact and placement rather than asking for an object somewhere in a vaguely selected background.

4. Generate and retain the returned candidates

Review the candidates produced by the selected operation. The recorded case returned four from one submission; that count is a property of that operation, not a universal promise for every Dreamina setting.

Do not discard a visibly flawed result before understanding the pattern. If all candidates alter the same label, a more attractive background does not solve the preservation problem. If only one has a convincing contact shadow, inspect its non-target areas before choosing it.

Download the candidate you intend to compare at full size. A small preview may hide changes to fine lettering, edge sharpness or surface texture.

Compare the result in three passes

Pass 1: did the requested change work?

Look only at the target first. Is the added object the right kind, size and number? Has the unwanted element actually disappeared? Did the corrected color reach the intended part without leaving obvious fragments?

Do not approve merely because something changed. A pumpkin in the selected region can satisfy the object-and-location request while still appearing to float. A removed sign can leave a distorted wall. The target must make sense in the scene.

Pass 2: does the transition belong?

Inspect the surrounding boundary. Look for halos, seams, smeared texture, broken lines, abrupt sharpness changes and inconsistent shadows. A local edit has to connect to the existing image as well as satisfy the prompt.

For removal, review what was reconstructed behind the object, including perspective and repeated texture. The object-removal and background-rebuilding guide develops that specific task. For a cutout or composite edge, the edge-repair guide addresses a different problem from generating an entirely new subject boundary.

Pass 3: did important non-target content drift?

Compare protected details with the baseline, even if they are far from the edited region. Start with a same-size side-by-side view. Then alternate the images or overlay them in a layer-capable editor, carefully keeping scale and placement aligned.

Check faces, letters, product outlines and small structural features at a useful zoom. Check the whole image at its final display size for shifts in tone, subject scale or composition. A change that is invisible in a thumbnail may be important in a large header or printed asset.

Use a simple decision note: target acceptable, boundary acceptable, protected details acceptable—or name the specific failure. This is more useful than saying the overall image is “90% good” when the remaining error is a mandatory logo.

Turn the comparison into an approval decision

For the bottle example, the next action depends on the brief. If the pumpkin is intended as a floating graphic accent, lack of ground contact may be acceptable. If it must look like a physical prop, its support and shadow need a coherent explanation. The same output can pass one use and fail another; decide that through the intended scene rather than through the presence of a pumpkin alone.

Do not repair every numerical difference simply because it exists. Concentrate on visible and factual requirements, and use a controlled pixel test only when numerical identity is itself part of the contract. That keeps review rigorous without turning harmless export variation into an endless editing task.

What a difference view can and cannot prove

In an editor that supports it, place the before and after images in aligned layers and use an opacity comparison or a difference view to locate changes. A difference view is a way to direct attention, not an automatic quality score.

First confirm both images have the same dimensions and alignment. If one has been resized or cropped, nearly every edge can differ even when the content is visually similar. Color-profile conversions, export settings and lossy compression can also complicate the comparison.

If you need an exact-pixel claim, keep those conditions controlled and examine the corresponding pixels outside a known mask. If you only need a visually faithful result, inspect whether the differences affect the approved content. These are different acceptance tests, and neither should be silently substituted for the other.

Also compare the final exported file. A working document may preserve an original layer, while a later resize or JPEG export changes its numerical values. “The source layer was kept” and “the delivered file has identical pixel values” describe different stages.

When another editing route is the right choice

Use an adjustment or original-pixel repair for a narrow factual change

If you only need to change brightness, color or a small blemish, an adjustment, healing or cloning operation may be more appropriate than generating a replacement region. Adobe's Spot Healing Brush guidance distinguishes small repairs on suitable surrounding texture from larger areas needing more control.

The choice follows the defect. Generating new material is useful when invention is acceptable; copying or retaining existing material is useful when faithful restoration matters.

Protect original artwork when exact details are mandatory

Keep the approved content on a retained layer and reveal the generated change only where it is needed. Inspect the mask edge and final composite. This gives you a way to control what is included rather than relying exclusively on a request to preserve it.

A layer-based method also requires judgment. A new background can disagree with the original subject's illumination or perspective. Protecting the source pixels solves retention, not every problem of scene integration.

Choose a documented selection workflow when you need more regional control

Midjourney's Editor documents selections, layers and regeneration of transparent areas. ChatGPT's image editor offers selections but notes that edits can extend beyond the highlighted region. Those are different documented behaviors, not evidence of a universal quality winner.

Choose the route that provides the needed operation and a workable comparison method. A single attractive example cannot establish that one application is always better at preserving every subject.

Fix the failure, then stop when the requirement is met

If the target is wrong, revise its description or selection. If the target is right but the boundary is wrong, repair the transition. If protected details changed, return to the baseline and use a route that retains them; repeatedly editing the drifted result can make the original harder to recover.

If the real task is adding space beyond the image, use the image-extension workflow. Expanding a canvas is different from revising an existing interior region, and the original boundary becomes part of the review.

Finally, export the approved version and inspect it in the intended layout. Retain the baseline, edit instruction, selected-area record and final file. Start a focused edit in Dreamina, but approve it only after the target, boundary and protected content all meet the actual brief.