Gemini edits the image itself rather than telling you how to edit it, and that changes how the prompt has to be written. The usual failure is not refusal. It is that the model quietly rebuilds something you never asked it to touch.
So every prompt below states what must stay identical before it says what changes. That single habit does more for edit quality than any amount of extra description, and it is why the prompts read as long as they do.
The fifty are grouped by what the edit targets: light and colour, removing and adding, background and scene, people, and a final group about keeping control across several turns. That last one matters more with Gemini than with most models, because you edit conversationally and it is easy to drift. For generating rather than editing, see Gemini prompts for image, and the parent Gemini prompts guide covers everything that is not an image.
How to Write an Edit Prompt Gemini Will Follow
Protect before you change. Name what must stay identical first. Adding do not alter anything else reads as redundant and measurably reduces the amount of collateral redrawing you have to undo.
One change per turn. Two edits in one instruction produce a compromise on both, and you lose the ability to tell which one caused the problem.
Ask what it changed. The single most useful follow-up available: list everything you changed, including anything I did not ask for. It catches the quiet crop and the smoothed skin.
Keep the original file. Conversational edits are not reversible the way a Lightroom slider is, and the version you liked is often two turns back.
A request looks like this:
Clean this photo up a bit.
The same ask as an edit prompt:
In the attached photograph, do not change the subject, the framing, the crop or the colour. Remove only the litter on the ground in the lower left, rebuilding the pavement texture to match. Then tell me everything you changed, including anything I did not ask for.
The second one is longer because most of it is protection rather than description. That is the part doing the work.
Light and Colour
The edits with the highest ratio of effect to risk, because none of them touches content. Working image to image, these are the safest place to start.
1. Change the light direction
Keep every object, position and framing in the attached image identical. Change the lighting so it comes from the upper left as hard late-afternoon sun, with defined shadows falling to the right. Warm the white balance slightly. Add nothing and remove nothing.
2. Fix a colour cast
Correct the green colour cast in the attached image without changing composition, sharpness or any content. Neutralise it in the midtones first, and tell me afterwards how much you shifted and in which direction.
3. Overcast to golden hour
Convert the light in the attached photograph from flat overcast to golden hour, with a low warm sun from behind the subject. Keep the subject, pose and every background object exactly as they are. Do not add lens flare.
4. Recover blown highlights
Recover detail in the blown-out sky in the attached image while leaving the rest of the frame untouched. Match the recovered area's grain and colour to the surrounding sky. Do not replace the sky with a new one.
5. Lift shadows without flattening
Open up the shadows in the attached image so detail becomes visible, but keep the deepest black point exactly where it is. Do not raise the overall exposure and do not touch the highlights.
6. Split-tone the image
Apply a split tone to the attached image: cool blue in the shadows, warm amber in the highlights, midtones left neutral. Keep every other aspect unchanged and tell me which zone you weighted most.
7. Match one image to another
Match the colour and contrast of the first attached image to the second, which is my reference. Change nothing about the content of the first image. List what you changed and by roughly how much.
8. Reduce a single colour
Desaturate only the red tones in the attached image, leaving every other colour at its current saturation. Do not convert anything to greyscale and do not alter luminance.
9. Add one practical light
Add a single warm lamp glow to the attached interior at the lower right, as though a lamp just outside frame had been switched on. Keep all existing light sources and shadows intact, and make the new shadows consistent with the old ones.
10. Black and white, done properly
Convert the attached image to black and white using channel weighting rather than desaturation, so the skin separates clearly from the background. Preserve the original tonal contrast, and tell me which channel you weighted.
Ask for the numbers afterwards. A model that reports how far it shifted something gives you a value you can reuse; one that just returns an image gives you a one-off.
Removing and Adding
The riskiest group, because the model has to invent something. Naming what is behind the removed object helps more than describing the object itself.
11. Remove one object
Remove [object] from the attached image. Reconstruct what is behind it so the repair is undetectable, matching existing texture, grain, shadow direction and colour. Leave everything else exactly as it is.
12. Remove a person from a crowd
Remove the person standing at [position] in the attached image. Rebuild the background and any partially hidden objects behind them plausibly. Do not alter anybody else in the frame, including their edges.
13. Remove the object, keep the shadow
Remove [object] from the attached image but keep the shadow it casts, as though the object were just outside frame. Everything else unchanged.
14. Insert with matched light
Insert [object] into the attached image at [position]. Match the existing light direction, shadow softness and the reflection characteristics of similar surfaces already in frame. Change nothing else at all.
15. Add something to empty space
The lower right third of the attached image is empty. Add [object] there at a scale consistent with the existing perspective, with a correct contact shadow. Do not reposition anything already in the frame.
16. Clean up the edges
Remove the wires, signage and clutter along the top edge of the attached image. Keep the sky texture and gradient continuous, and leave the building silhouettes exactly as they are.
17. Remove reflections in glass
Remove the reflections on the glass in the attached image so what is behind it becomes visible. Keep the frame, the glass edges and everything outside the glass untouched.
18. Fill a missing area
The attached scan has a missing corner. Reconstruct it from the surrounding content, matching the paper texture, grain and any pattern that continues into that area. Do not retouch anything else.
19. Replace one element only
Replace the vehicle in the attached image with a bicycle at the same position and scale. Keep the road surface, shadows, lighting and everybody in the frame exactly as they are.
20. Add weather
Add light falling snow to the attached street scene, with a thin accumulation on horizontal surfaces only. Cool the white balance slightly to match. Do not change the light direction or add any new light source.
Check the repair at full size, not on a phone. Reconstructed areas usually look convincing until you zoom in, and that is exactly where a client will look.
Where the subject stays and everything behind it moves. The edge detail on hair and fabric is the first thing to check on any of these.
21. Background replacement
Keep the subject in the attached image exactly as is, including the edge detail in the hair. Replace the background with a seamless mid-grey studio backdrop and adjust the subject's edge lighting to match a single softbox from the upper left. Do not alter the pose, skin tone or clothing.
22. Blur the background convincingly
Reduce the depth of field in the attached image. Keep [subject] sharp and blur the background progressively by distance, with natural circular bokeh matching an 85mm lens at f/1.8. Preserve the original colour grade.
23. Extend the frame
Extend the attached image outward to a [ratio] frame. Generate plausible content continuing the existing scene, matching grain, colour grade and perspective exactly. Do not alter or rescale the original area, and keep the horizon consistent.
24. Change the season
Convert the attached park scene from summer to winter. Keep the composition, paths and buildings unchanged. Replace foliage with bare branches, add fresh untouched snow on horizontal surfaces, cool the white balance, and flatten the light to overcast.
25. Change the time of day
Convert the attached exterior from midday to blue hour. Keep every structure and object in place. Add practical window lights in the buildings and adjust the shadows to match the new light level.
26. Tidy the room, not the person
In the attached image, remove the clutter from the background room while keeping the subject and everything they are touching exactly as is. The room should still look lived in rather than staged.
27. Swap the setting
Place the subject from the attached image into [new setting], matching perspective, eye level and light direction to the new environment. Keep the subject's pose, clothing and expression identical.
28. Simplify a busy background
Reduce the visual complexity of the background in the attached image by removing competing detail, while keeping the location recognisable. Do not blur it and do not change the subject.
29. Fix a crooked horizon
Level the horizon in the attached image and fill any resulting gaps at the edges by extending the existing content. Do not crop into the subject.
30. Add depth to a flat scene
Add atmospheric haze to the far distance in the attached landscape so the depth reads more clearly. Keep the foreground and midground exactly as they are and do not change the colour of the sky.
Always name the edge detail you want protected. Hair, fur and fabric edges are where a background swap gives itself away, and asking explicitly is what stops it.
People and Portraits
The group where restraint matters most, because a face is the thing a viewer checks hardest. For keeping the same person consistent across several images, our Nano Banana prompts go deeper, and Nano Banana Pro is the tier worth using for it.
31. Retouch with restraint
In the attached portrait, remove temporary blemishes only. Leave texture, pores, lines and all permanent features intact. Do not smooth the skin, do not change the face shape, and do not alter the eyes.
32. Even out the light on a face
Reduce the harsh shadow on the right side of the face in the attached portrait, as though a reflector had been added. Keep the light direction, the catchlights and the background exactly as they are.
33. Open closed eyes
In the attached group photograph, the second person from the left has their eyes closed. Open them to match their expression and the direction everybody else is looking. Change nothing else about their face.
34. Change the outfit only
Change the subject's t-shirt in the attached portrait to a charcoal linen blazer over a white shirt, with natural drape and a correct shoulder fit. Keep the face, pose, hair and background completely unchanged, and match the fabric lighting to the existing key light.
35. Same person, new setting
Using the person in the attached reference, generate three more images of them in different settings. Keep the face, hair length, build and clothing exactly as in the reference. Change only the environment and the lighting.
36. Fix the skin tone, not the person
Correct the orange skin tone in the attached portrait to something natural for this person, without changing the background colour or the overall exposure. Tell me what you shifted and by how much.
37. Remove somebody from a portrait
Remove the person on the right from the attached photograph, rebuilding the background behind them. Do not alter the remaining subject in any way, including their edges and the shadow they cast.
38. Add catchlights
Add subtle catchlights to the eyes in the attached portrait, consistent with a single large light source from the upper left. Change nothing else about the face or the lighting.
39. A small posture correction
Slightly correct the subject's posture in the attached image so the shoulders are level. Keep the face, clothing and background identical, and keep the correction small enough to look unedited.
40. Match two portraits
These two portraits of the same person were shot in different light. Match the second to the first in colour, contrast and skin tone. Do not change either person's expression, pose or framing.
Write do not smooth the skin into every portrait prompt. Smoothing is the default failure, it is applied without being asked for, and it is the hardest thing to undo once the texture has gone.
Iterating and Keeping Control
This group has no equivalent in a slider-based editor, and it is the part people get wrong. Because you edit conversationally, drift accumulates quietly across turns. These ten are for noticing it. The same ideas applied to a text model are in our ChatGPT prompts for photo editing.
41. One change per turn
I am going to ask for several edits to this image, one at a time. After each, change only what I asked for and confirm in one line what you altered. Do not make additional improvements you think would help.
42. Ask what changed
You have just edited this image. List everything you changed, including anything I did not ask for. Be specific about crops, colour shifts and any smoothing.
43. Protect a region explicitly
Edit only the area of the attached image to the left of the doorway. Treat everything to the right as locked, including its edges. If the edit would affect the locked area, tell me instead of doing it.
44. Go back a step
Return to the version before the last edit and apply this instead: [new instruction]. Keep everything that was working in the earlier version.
45. Compare two versions
Here are two versions of the same edit. Describe the differences precisely, and tell me which better preserves the original image's character rather than which looks more impressive.
46. Stop before over-editing
Continue improving this image, but stop as soon as further changes would be a matter of taste rather than correction. Tell me when you reach that point and what the remaining choices are.
47. Diagnose before editing
Before you change anything in the attached image, tell me the three things most responsible for it not working, in order, and which one to address first. Do not edit yet.
48. Ask for the risk first
Before applying this edit, tell me what is most likely to go wrong with it on this particular image, and what I should check in the result.
49. Reproduce the edit elsewhere
Describe the edit you just applied precisely enough that it could be reproduced on a different photograph, including what should change if that photograph is exposed differently.
50. Keep the original
Apply this edit and describe it, but confirm that the version I uploaded is unchanged, and tell me how to name the two files so I do not lose the original.
Prompt 42 is the one to memorise. Asking what else changed costs a single message and catches the quiet alterations that are invisible on a phone screen and obvious in print.
Using These Prompts in Chat Smith
Editing is where having several models in one place earns its keep, because they fail differently and you cannot tell in advance which handles your kind of image. Gemini 3.5 Flash and the Nano Banana tiers behave differently on the same file, particularly on faces and on reconstructed backgrounds. Running one edit through two of them takes a minute and settles the question.
If you would rather generate a new image than edit an existing one, work straight from an AI image generator. Chat Smith is free to try, and the conversation keeps your earlier versions, which is what makes the one-change-per-turn discipline practical rather than theoretical.
Gemini prompts for photo editing are written instructions that tell the model how to change a photo you already have rather than generate a new one. A workable prompt states what must stay untouched, what should change, and the finish you want, such as a cleaner background or warmer light.
The Chat Smith Editorial Team is a group of AI enthusiasts, researchers, and content creators passionate about making artificial intelligence more accessible and practical. Through the Chat Smith blog, we share the latest AI trends, tool reviews, industry insights, and actionable guides to help individuals and businesses get more value from AI. Our mission is simple: deliver clear, reliable, and easy-to-understand content that helps readers stay informed, productive, and ahead in the fast-moving world of AI.
Share this article
Related Articles
Level Up Your Work,One Click Away!
Everything you need to push projects forward is right at your fingertips.
