Replace an object in place
Paint a product, prop, or distracting item and describe the substitute. The unpainted scene stays as context so the new object can inherit tabletop texture, shadow direction, and camera angle.
Brush-selected region fill for photos
Generative Fill
Use Generative Fill to upload a photo, paint the region to change, describe the fill, generate a spatially masked edit, compare it with the source, and download the result.
JPG, PNG, WebP Brush and eraser 20 credits
Masked edit
Fill a Painted Region Online
Paint the region to change, then write a fill prompt. Generation uses a 1536 px working canvas.
JPG, PNG, or WebP20 MBWorking size ≤ 1536 px
Example input
Product replacement
Upload a JPG, PNG, or WebP image first.
What Is Generative Fill?
Generative Fill is a masked in-frame edit: you paint what should change, describe the replacement, and generate a new photograph in which that region is rewritten.
Unlike a whole-image instruction, Generative Fill limits the request to a painted mask. The unpainted pixels supply lighting, perspective, and material clues, so the new content has to meet a real surrounding scene instead of inventing an entirely new composition.
That is still a generative proposal, not a recovery of hidden pixels. If a bottle, sky, or table gap is covered, the model cannot know what a camera would have seen there. Inspect the filled boundary before you publish, and use a different tool when the job is not an in-frame region edit.
Choose the AI Image Editor for a prompt-led change across the full frame, Uncrop when you need more canvas outside the photo, or Background Remover when the subject should be isolated as a transparent PNG.
What Generative Fill Can Do
The workbench is built for local, in-frame changes. It is strongest when one region should become something else and the rest of the photograph is already the scene you want to keep.
Rebuild sky or background patches
Cover a blown-out sky, a busy wall, or a slice of backdrop and ask for a continuation that agrees with the remaining horizon, color temperature, and depth of field.
Add an object into a gap
Paint empty table space, a shelf, or a floor patch and describe the object that should appear there. Keep the mask inside the gap so surrounding furniture and architecture remain the reference.
Leave unpainted pixels in place
The fill is spatially masked: white painted pixels are the edit region, and the rest of the working canvas is preserved as the surrounding photograph rather than rewritten as a whole new image.
It will not enlarge the frame, strip a background to transparency, or reconstruct a watermarked overlay as its primary job. Those remain Uncrop, Background Remover, and the watermark workflow mapped as Object Remover at Remove Watermark.
How to Use Generative Fill in Four Steps
The focused path is upload, paint, describe, then review. Signed-out visitors can load an image and paint locally; sign-in is requested only when Generate runs.
- 01
Upload one source image
Choose a JPG, PNG, or WebP up to 20 MB. The source is decoded locally first. If the longest side is larger than 1536 px, it is downscaled once to a working canvas that stays fixed for the session.
- 02
Paint the region to change
Use the brush to mark what should be replaced. Resize the brush, erase extra strokes, zoom, pan, undo, and redo until the overlay covers the object, patch of sky, or interior gap you want to change.
- 03
Describe the fill
Write a 3–1000 character prompt that names what belongs in the painted area and what nearby lighting, material, and perspective should match. Suggestion chips are a starting point, not a required formula.
- 04
Generate, compare, and download
Generation costs 20 credits. The result replaces the stage in place. Hold to compare with the original, inspect the painted boundary, then download or regenerate with the same prompt if the fill still needs work.
What to Paint and How to Prompt
A useful mask covers the intended region without swallowing nearby faces, logos, or architecture. A useful prompt names the fill and the lighting or material it should match.
Cover the whole target, then stop
Paint every pixel of the object or area you want changed, including a thin margin around complex photographic edges. Leaving holes inside the mask often leaves fragments of the original object. Painting far beyond the target asks the model to invent more scene than you intended.
Follow the real boundary
Zoom in on hair, glass, lettering, and furniture corners. A mask that follows the visible edge is easier to blend than a loose scribble. Use the eraser when a stroke spills onto a face, logo, or product silhouette that must stay.
Describe the fill, not a new story
Short, concrete prompts work better than a list of unrelated changes. “Replace the painted bottle with a matte ceramic vase, same scale, warm left window light” is easier to honor than a request that also restyles the room, the clothes, and the camera.
Match light, material, and scale
Name the material, the time of day, and how large the new object should sit relative to nearby items. If the painted region is a sky, mention cloud density and color rather than a different landscape. If it is a tabletop, mention wood grain, marble, or linen so the fill can meet the unpainted surface.
Generative Fill Before and After
These three cases keep the same composition inside each pair: a product swap, a sky rebuild, and an object added to an interior. They illustrate the mask-and-prompt job rather than a whole-image restyle.
Object replacement
Swap a product in a tabletop scene
Paint the original bottle or prop and describe a replacement that keeps the same set, reflections, and crop. This is the job when the surrounding campaign photograph is already approved.
Sky and backdrop
Rebuild a sky or background patch
Cover the washed-out sky or a distracting backdrop and ask for clouds, dusk color, or a quieter wall. The buildings, horizon, and foreground stay outside the mask so the new atmosphere has to meet them.
Interior add-in
Add an object to an interior
Paint a gap on a table, shelf, or floor and describe a vase, lamp, or book that belongs in that space. The room geometry stays unpainted so the added object has to sit in the existing perspective.
Honest Limits of a Painted Fill
A painted fill can look continuous and still be wrong at the seam. Knowing where the tool struggles prevents over-promising on edges, lettering, and faces.
What it can do
- Repaint a selected region while using the unpainted photo as context.
- Replace, rebuild, or add objects when the mask and prompt agree with nearby light and scale.
- Let you erase, zoom, and undo the mask before spending credits.
- Compare the filled result with the original on the same stage.
What it cannot promise
- Pixel-perfect blending along complex photographic edges such as hair, glass, foliage, or motion blur.
- Reliable reconstruction of small text, fine logos, or tiny labels inside or beside the mask.
- Faithful identity on small faces, hands, or jewelry that sit on the painted boundary.
- Recovery of the real hidden scene, or a whole-image restyle from a few brush strokes.
When an edge, a line of small text, or a face looks off, tighten the mask and repaint rather than accepting the first proposal. If the job is really a full-frame change, switch to the AI Image Editor. If the frame itself is too tight, use Uncrop. For broader model options, compare the GPT Image 2 guide.
Credits and Entitlements
One generation is one reserved 20-credit task. Painting locally is free. The wall for sign-in and balance appears when you generate, which is as late as the site-wide auth policy allows.
What 20 credits buys
Each Generate press reserves 20 credits for one fill at the working canvas size. There is a single credit tier for this tool. The button label states the cost before you run it. Insufficient credits link to pricing instead of silently failing.
Sign-in, refunds, and retries
Signed-out visitors can upload and paint. Generate asks for sign-in. Failures, cancellations, and duplicate callbacks refund the reservation once. A retry mints a new request so you are not charged twice for the same press. Completed results appear in My Creations.
Generative Fill vs Object Remover vs Uncrop
Fill repaints inside a painted region. Object Remover on this site is the watermark-removal workflow: it erases an overlay without adding a requested subject. Uncrop extends the photograph outside the current frame.
| Decision | Generative Fill | Object Remover | Uncrop |
|---|---|---|---|
| What changes | Repaints only the painted region and leaves unpainted pixels as the surrounding photograph | Removes a visible overlay or object by reconstructing covered pixels, without adding a new requested subject | Adds newly generated scene area outside the current frame |
| How you steer it | Brush and eraser plus a fill prompt for the masked area | A location hint and prompt-led reconstruction of the marked overlay | A target aspect ratio and optional scene guidance for the new edges |
| Best when | One in-frame region should become something else while the rest of the photo stays | A watermark or similar overlay should disappear from an image you are allowed to edit | The subject is useful but the crop is too tight for the destination layout |
| What it is not | Not a whole-image restyle, not an uncrop, and not a recovery of hidden real content | Not a brush-selected add or replace tool | Not an in-frame object editor |
Choose Generative Fill
When one in-frame region should be added, replaced, or rebuilt and the rest of the photo should stay.
Choose Object Remover
When a watermark or similar overlay should disappear from an authorized image. That workflow lives at Remove Watermark.
Choose Uncrop
When the subject is right and you need a wider or taller frame rather than a painted in-frame change. Open Uncrop.
Generative Fill FAQ
Straight answers about the mask, unpainted pixels, supported files, credit use, hidden-content limits, and when to repaint.
What is Generative Fill?
Generative Fill is a spatially masked image edit. You upload one photo, paint the region that should change, describe what belongs there, and generate a new image in which the painted area is rewritten and the unpainted area remains the surrounding photograph.
Does the fill change unpainted pixels?
The workflow is designed so only the painted region is the edit target. Treat that as a practical boundary, not a laboratory guarantee: inspect identity, logos, small text, and edges after every run, especially where the mask meets complex photographic edges.
Which files can I upload?
The workbench accepts JPG, JPEG, PNG, and WebP images up to 20 MB. Use the sharpest source you have. Images whose longest side exceeds 1536 px are downscaled once to a working canvas of at most 1536 px.
How many credits does a fill use?
Each generation costs 20 credits. The primary action shows that cost before you run it. Sign-in and available credits are checked when you generate. Local painting does not spend credits. Failed, canceled, and duplicate-settled tasks are refunded through the existing credit machinery.
Can Generative Fill recover what was behind an object?
No. The fill invents a visually plausible replacement from the prompt and the unpainted context. It does not restore a factual scene that the camera never recorded, and it should not be used as evidence of what used to sit in that region.
When should I repaint and try again?
Repaint when the mask missed part of the object, spilled onto a face or logo, or left a hole. Rewrite the prompt when the fill ignored material, scale, or lighting clues. Complex edges, small text, and small faces are the most common places a second, tighter mask helps.
Paint a region and describe the fill
Upload the best available source, cover only the area that should change, and write one concrete replacement.
Start a fill