Start with one clear instruction, inspect the result beside the original, then mark up to five areas for a more focused second pass. Every successful edit becomes a version you can revisit.
1
Upload photo
JPG, PNG or WEBP · up to 10 MB
Drop a photo hereor click to choose
Use a clear photo and one main change. Avoid explicit or graphic content. Do not ask the edit to add text, logos, or watermarks.
2
Describe the change
Say what to change and what to keep.
Upload a photo first to enable this button.
Charged only after a result is saved. Failed edits use 0 credits.
One prompt. A visible difference.
Before
After
Demo prompt Open the corner shop with cobalt, coral, and amber light. Keep the building and camera angle.
✓Change the whole image
✓Compare before and after
✓Mark up to 5 areas to refine
Reviewed examples
What a specific prompt can change
Each example names both the requested change and the details that should be preserved.
Before
After
Light a rainy storefront for a night campaign
Prompt “Open the corner storefront as a contemporary art shop with cobalt, coral, and amber light. Strengthen the wet-street reflections. Keep the camera angle and architecture unchanged.”
Prompt “Remove the red cable, cardboard box, packing material, floor lamp, and magazines. Reconstruct the rug, wall, baseboard, and floor. Keep the chair, table, window, and light unchanged.”
Prompt “Keep the road curve, cabin, cliff, coastline, and camera angle. Restyle the scene as a hand-inked Western relief print with a limited cobalt, coral, cream, green, and yellow palette.”
Plan an AI image edit that survives a careful review
A useful edit is more than a visually impressive first result. The source has details that carry meaning: a product silhouette, a person's identity, the angle of a room, a sign, or a specific reflection. The workflows below show how to separate the requested change from the evidence that should remain stable, then use the comparison and version tools to judge the result.
Case 01
Clean a real scene without making it look rebuilt
For the reading-nook example, the unwanted cable, box, packing material, lamp, and magazines sit on different surfaces. Naming the rug, floor, wall, and baseboard tells the editor what must be reconstructed behind each object. It also protects the chair, table, plants, window, and daylight from becoming accidental collateral changes.
PromptRemove the red cable, box, packing material, floor lamp, and magazines. Reconstruct the rug, wall, baseboard, and floor. Keep the chair, table, plants, window, perspective, and light unchanged.Use
Review: Compare the rug edge, floorboard direction, baseboard line, plant leaves, and chair legs. If one seam is wrong, mark that seam for a child version instead of asking the next prompt to clean the entire room again.
Case 02
Change atmosphere while protecting architecture
A night-scene edit can easily redesign a building when the actual goal is lighting. Treat the storefront, windows, road markings, curb, and camera angle as a preservation list. Then describe the light sources, color palette, and wet-street reflection as the permitted change. This gives the model a clearer boundary than a broad request to make the image cinematic.
PromptOpen the corner storefront as a contemporary art shop with cobalt, coral, and amber light. Strengthen wet-street reflections. Keep the building, windows, road, curb, and camera angle unchanged.Use
Review: Drag the comparison line across the roofline and window grid before looking at color. Check that road perspective and curb height still match. Refine a single sign or reflection only after the large geometry passes.
Case 03
Create product art direction without redesigning the product
Product edits fail commercially when the generated scene is attractive but changes proportions, hardware, materials, or labels. Put the product identity first. In the chair example, the cobalt chair, camera position, silhouette, seams, and legs are protected while the plain studio becomes a campaign set with a new platform, wall color, and harder light.
PromptKeep the cobalt chair, silhouette, seams, legs, and camera position unchanged. Replace the plain studio with a coral architectural set, an acid-yellow platform, directional light, and hard editorial shadows.Use
Review: Inspect the outer silhouette, leg count, seat angle, seam placement, contact shadow, and any label at full size. A traditional editor remains safer when a regulated package or exact trademark must be pixel-identical.
Case 04
Repair an old or damaged image without erasing its character
Restoration is an editing task when the goal is to remove scratches, dust, tears, or fading while keeping the people, objects, framing, and period character recognizable. Describe the damage and the photographic qualities that should continue behind it. Avoid asking the model to modernize or beautify the complete image unless that is a separate, deliberate version.
PromptRemove the visible scratches, dust marks, and small paper tear. Reconstruct the affected clothing and wall texture from nearby detail. Preserve every face, expression, pose, object, crop, monochrome tone, film grain, and original light.Use
Review: Compare facial proportions, hands, clothing seams, jewelry, furniture, and edge grain. Restoration cannot recover hidden truth; it generates a plausible repair. Keep the source and label the result clearly when historical accuracy matters.
Choose the right amount of AI
One editor does not make every job the same.
Use a whole-image edit
Choose the broad editor when the new request affects the scene, atmosphere, clothing, style, or several connected surfaces. Keep the instruction focused on one visual goal and include a preservation list. Save the first acceptable result before experimenting with a riskier direction.
Use marked refinement
Choose marked refinement after a usable result exists and one to five local details still need work. Give every marker one job. The markers guide attention; they do not create a hard mask.
Use a manual editor
Choose a conventional mask or retouching tool when untouched pixels must remain mathematically identical, text must be exact, or the image is used for legal, medical, identity, or regulated product evidence.
Before download
Review the image, not just the effect
The checklist turns an attractive result into a reviewable asset. Use it on every saved version, especially when the image contains identity, text, products, architecture, transparent material, or commercial claims.
The prompt names one primary change in observable language.
Protected people, products, text, geometry, and camera details are listed explicitly.
The result is checked against the source with the before-and-after control.
Faces, hands, labels, small text, reflections, and transparent edges are reviewed at full size.
A local problem becomes a marked child version instead of another broad rewrite.
The downloaded result is treated as generated output and receives a final human review.
How to use it
One change first, detail second
1
Start with a clear source
Upload a JPG, PNG, or WEBP under 10 MB. Keep the subject and important edges visible.
2
Name change and preservation
Say what should change, then list the face, pose, product, text, or composition that should stay.
3
Inspect and refine
Compare before and after. Mark up to five areas for a more focused follow-up version.
Good fits
When to use this workflow
Precise follow-up edits
Mark the part that still needs attention instead of restarting from the original.
Consistent creative versions
Keep a parent-and-child history so each experiment stays connected to its source.
Everyday photo cleanup
Remove distractions, repair damage, adjust styling, or change atmosphere with plain language.
Yes. Generate a first result, open Refine marked areas, add up to five markers, and give each marker a separate instruction.
Will the rest of the photo remain identical?+
ImageRework asks the model to preserve unmarked areas, but the first version does not promise strict pixel locking. Always compare details that must remain exact.
When are credits charged?+
A successful whole-image edit or refinement uses 3 credits. Failed, timed-out, and duplicate submissions use 0 credits.
Can I return to an older result?+
Yes. Each edit is saved as a version in the same project so you can inspect and continue from an earlier result.
