How to Change One Part of an Image with AI Without Redoing the Whole Photo
A good local edit starts before you draw a marker. First establish a usable base result, then describe the small change and the surrounding details that must stay stable.
Traditional image editors use masks to define exact pixels. A prompt-based AI editor works differently: a marker tells the model where to focus, while the source image and your words explain what belongs there. This makes the workflow fast and flexible, but it also means a marked region is guidance rather than a hard pixel lock.
Use a two-pass workflow
1. Make the broad edit first
Upload the source and ask for the main result in one sentence. If the goal is a professional portrait, change the setting first while explicitly preserving the face, clothing, pose, and hair edges. Inspect the whole image before moving to smaller details.
“Replace the background with a softly lit modern office. Keep the person, face, hair edges, clothing, pose, and camera angle unchanged. Match the light naturally.”
2. Mark only the unresolved area
Open Refine marked areas on the saved result. Add a marker near the problem. ImageRework supports up to five markers, but fewer is usually clearer. Give every marker one concrete instruction. Use the optional whole-image direction for preservation, not for adding another large transformation.
“Clean the flyaway hair boundary here and make the transition into the background natural.”
Whole-image direction“Keep the face, hairstyle, clothing, pose, office setting, and overall lighting unchanged.”
What makes a marker instruction useful
- Name the location through the marker. The circle already indicates “where,” so the text can concentrate on the desired correction.
- Describe the visible outcome. “Clean natural hair edge” is clearer than “fix this.”
- Separate unrelated changes. A collar correction and a background correction should use different markers and instructions.
- Repeat important preservation rules globally. If identity or product shape is critical, state it again.
Choose the base version before you mark anything
Local refinement is most effective when the current image is already good enough to protect. Do not place markers on a result whose composition, identity, product shape, or overall art direction is still wrong. A marker narrows the next request; it does not repair a weak foundation. Compare the source and current result first, then decide whether the next problem is truly local.
Use a whole-image edit again when several connected parts need the same large correction. Use marked refinement when you can point to one to five specific areas and give each one a separate outcome. If the requested change affects the entire medium, season, or visual language, an image-to-image workflow is usually clearer.
Three focused editing patterns
Repair a reconstruction seam after object removal
Removing an object means generating what might have been hidden behind it. In the reading-nook example, a cable crosses a rug and floor, while boxes and packing material cover additional surfaces. The broad prompt names the objects and the rug, wall, baseboard, and floor that need rebuilding. After generation, one edge may still look inconsistent even when the room is usable.
“Continue the rug edge through this area with the same weave, border width, perspective, and shadow. Do not change the nearby chair leg or wooden floor.”
Whole-image direction“Keep all furniture, plants, window light, room geometry, and the rest of the cleaned result unchanged.”
The review should concentrate on the rug border and the neighboring chair leg, but it should not stop there. Drag across the complete image and confirm that the model did not reintroduce a removed object or alter the window, plant, or furniture elsewhere.
Correct one product edge in a campaign image
A generated campaign set can look excellent while a product seam, leg, label, or contact shadow is wrong. Mark the smallest meaningful area. The instruction should name both the commercial feature and the surfaces it meets. “Fix the chair” is too broad; “restore the straight rear leg and its contact shadow” creates a reviewable target.
“Restore the chair's straight rear leg here using the source shape and cobalt material. Rebuild the contact shadow on the yellow platform. Keep the seat, other legs, set, camera angle, and lighting unchanged.”
When exact product geometry or a trademark must remain pixel-identical, use a conventional retouching or compositing workflow for the final correction. Generative refinement is useful for art direction and cleanup, but it should not be described as certified product reconstruction.
Reduce one reflection without flattening the scene
Reflections connect an object, light source, viewing angle, and surface. A marker can identify the problem, but the text still needs to explain the relationship. For a wet street, specify which colored reflection should be shorter or softer and keep the curb, lane markings, storefront, and other reflections stable.
“Shorten and soften the coral reflection in this area so it follows the road perspective and wet asphalt texture. Keep the blue and amber reflections, curb, road markings, and storefront unchanged.”
How many markers should you use?
ImageRework supports up to five markers, but the maximum is not the target. One marker with one concrete job is easiest to judge. Use two or three when the areas are separate but share the same stable base image. For example, a portrait may need one marker for the fringe, one for a collar, and one for a background edge. Give each marker its own instruction rather than placing three circles around the image and describing everything in one paragraph.
If five markers still do not cover the request, decide whether the change is actually local. Several related corrections may be clearer as a new whole-image edit. Conversely, do not repeat the same broad preservation paragraph inside every marker. Put shared rules—identity, composition, product, lighting, text, or style—into the whole-image direction.
Review local edits in three passes
- Target pass: zoom into every marker and confirm that the requested visible outcome exists. Look for seams, repeated textures, broken lines, invented text, or unnatural transitions.
- Context pass: inspect the objects and surfaces immediately around the marker. A correction may be locally attractive but inconsistent with perspective, shadow, material, or anatomy.
- Whole-image pass: compare the entire result with its parent version. Confirm that identity, architecture, product shape, color, lighting, and other protected details remain acceptable.
Download only after all three passes. A before-and-after slider is useful because human attention is drawn to the edited area and can miss a smaller change elsewhere. Moving the comparison line across the complete frame forces a broader check.
When a second local pass is the wrong answer
Stop refining when the model repeatedly changes a detail that must remain exact, when the source lacks enough information to reconstruct hidden content, or when a transparent or reflective relationship depends on a large part of the scene. Repeated generations can consume time without creating new evidence. Return to the parent version, simplify the prompt, use a better source image, or finish with a traditional masked editor.
A focused edit is successful when the corrected area improves and the rest of the image still passes review—not merely when the marker itself looks different.
A compact marked-refinement brief
Before submitting, read the marker list once as if you had not seen the image. Each item should identify a visible result, and the whole-image direction should protect the approved version. A compact brief usually contains: “Marker 1: correct this edge and match this surface. Marker 2: restore this named object feature. Whole image: preserve identity, composition, camera, light, and every approved element.” If a marker depends on a long explanation of several distant objects, the task is probably too broad for a local pass.
Keep the parent version open during review. The goal is not to prove that the new image changed; it is to verify that the change solved the stated problem with less collateral damage than a new whole-image generation. Record which marker passed.
Where focused editing can still fail
Very small text, transparent materials, fingers crossing an object, and reflections often involve relationships outside the circle. The model may adjust nearby pixels to make the correction believable. If the surrounding content must remain exact for legal, medical, or product-accuracy reasons, use a traditional masked editor for the final production step.
Use marked refinement for a faster, more focused second pass—not as a promise that every unmarked pixel is locked.
Keep the version that worked
Every successful ImageRework refinement becomes a child version in the same project. If the second pass improves one detail but harms another, return to the earlier result and start a different branch. This is safer than repeatedly overwriting the only usable image.