AI Picture Prompts for New Images and Photo Edits

The first result for ai picture prompts is a Reddit thread from someone trying to modify photos from their camera and generate new ones, who says that whatever prompt they give, it just isn't working out. Below that are the banks. Two hundred and sixty prompts here, 1,895 there, fifty on the next page, none of which would have helped, because the poster's problem is structural and a bank is a pile of strings.

Two shapes, and they're different. A prompt for a picture from nothing has four slots, subject, setting with its light, style, and framing. A prompt for a picture you already have has three, the reference, one change, and a keep clause that says what must not move. My bet is that most failed photo edits are missing the keep clause and most generic new images are missing the light.

Both shapes are below, six full prompts each, then the fixes sorted by what went wrong.

Two Shapes, Side By Side

The row that trips people is the second one. Describing the light in an edit prompt tells the model to relight the photo, which it then does, and the edit comes back looking like a different day. Google's editing template for its image model puts it plainly. Provide an image and describe your change, and "the model will match the original image's style, lighting, and perspective". So the lighting belongs to the photo. Leave it alone unless the change is the light.

Six Prompts For A Picture From Nothing

Each is a working prompt, then a note on what the slots are doing. The depth on any single style lives in the sibling posts, so these stay short.

A woman in her sixties at a market stall, laughing at something off frame, holding a paper bag of oranges. Late afternoon sun from the left, long shadows across the stall. Shot on a 50mm lens at f/2, slight film grain, muted colour. 4:3.

Subject with an action and a prop, which stops the default pose. One light source with a direction. A lens and an aperture instead of the word "professional". A ratio.

A ceramic mug, matte sage green, on a raw oak board. Single softbox from above and slightly behind, so the rim catches light and the front falls into soft shadow. Plain grey paper backdrop. Product photography, no props, no text. 1:1.

Product shots fail by addition. A prop appears, or a second shadow, or a word on the mug. So "no props, no text" is doing real work, and the light description names where the highlight lands rather than asking for "studio lighting", which is the difference between a catalogue image and a render that looks like every other render on the same page.

A narrow street in a hill town after rain, empty, one lit doorway at the far end. Blue hour, the sky just brighter than the stone. Wet cobbles reflecting the doorway. Painterly but photographic, soft edges, cool palette with one warm point. 9:16.

This is the "aesthetic" search made specific, and the shape I'd hand to anyone who searched for it. Aesthetic prompts are usually a mood word and nothing else; this one puts the mood into a time of day, a colour relationship, and one warm point in a cool frame, all things the model can actually place.

A bowl of ramen, chopsticks lifting noodles mid-frame, steam rising. Overhead window light from the right, shallow depth of field, the bowl sharp and the table soft. Dark wooden table, one folded napkin. Food photography, natural colour, no garnish added. 4:5.

Food defaults to over-garnished and over-saturated, so both are forbidden by name. The chopsticks in motion give the image a moment rather than a display. Short prompt. It's enough.

A greyhound asleep on a striped sofa, curled tight, one ear folded back. Morning light through a window behind the sofa, backlit so the ear glows. Domestic, slightly untidy room, a mug on the floor. Candid photo, 35mm, warm tones. 3:2.

Pets read as stock unless something is particular, and I'd pick the particular before the breed. The folded ear, the mug on the floor, the untidiness. Backlighting through a thin ear is a real photographic effect the model reproduces well.

A poster for a small-town film festival. Bold two-word title "NIGHT REEL" at the top, hand-painted lettering. Below it, a stylised projector beam over a lake at night, flat colour, 1960s screen-print style, three inks. Portrait, 2:3.

I keep text short because text is where these models still fail most often. Two words, capitalised, in a style that forgives wobbly letterforms. Screen-print style also limits the palette, which makes the whole thing cohere. For the full clause-by-clause treatment of generation prompts like these, ai image prompt examples walks through the edits that fixed them.

Six Prompts For A Photo You Already Have

This is the half the banks skip and the Reddit poster needed. The shape is reference, one change, keep clause, and Google's own inpainting template is the cleanest statement of it I've found. It reads "Using the provided image, change only the [specific element] to [new element/description]. Keep everything else in the image exactly the same, preserving the original style, lighting, and composition." Every prompt below is a variation on that sentence.

Using the provided image, change only my jacket to a dark green wool overcoat, knee length, buttoned. Keep everything else exactly the same: my face, hair, pose, hands, the background, and the lighting. Match the coat's shadows to the existing light direction.

One change, stated with enough detail to be one change (length, colour, fastening). The keep list names the things that drift. The last sentence tells the model that the new object has to obey the old light, which it tends to do when asked and skip when not, and a coat lit from a different direction than the face above it is the commonest tell in a clothing swap.

Using the provided image, remove the person in the red jacket on the left. Fill the space with a continuation of the existing background, matching the paving and the wall. Keep everyone else, the framing, and the lighting exactly as they are. Do not add anything new.

Removals fail by addition. The model fills a gap with something interesting unless told, twice, that the gap is boring, which is why "continuation of the existing background" and "do not add anything new" both appear in a prompt that could in theory have said either.

Transform the provided photograph into a loose gouache painting, visible brush marks, slightly simplified shapes. Preserve the composition, the position and expression of the people, and the direction of the light. Do not change the crop.

The style transfer case, where the look is the change. Google's template for this opens with transforming the provided photograph into the style of an artist or art style and then, like the others, moves on to what should be preserved. The preserve list is doing the same job as the keep clause, just with the style released.

Using the provided image, change only the sky to an overcast, evenly lit grey. Adjust the ground shadows to match a diffuse sky, softer and less directional, but keep everything else, including the buildings, the people, and their positions, exactly the same.

The one case where you do touch the light, because the change is the light. Note that it says what should follow from the change (softer shadows) so the model doesn't leave hard noon shadows under a grey sky. That contradiction is the tell I check first in a sky swap.

Using the provided image of my dog, place the dog on a rocky beach at low tide, late afternoon. Keep the dog's exact pose, proportions, fur colour, and expression. Change the background and the lighting to suit the new location, with a warm low sun from the right.

Background replacement is the inverse of the jacket prompt. The subject is the keep clause and the scene is the change, and because the scene changes, the light must be restated. This is the one time in the edit set that a light direction appears, and it appears because it has to.

Using the two provided images, put the person from the first image into the scene from the second, standing to the right of the bench. Match the scale to the bench and the lighting to the second image. Keep the person's face, clothing, and pose from the first image unchanged.

Multiple references, which Google's docs allow with a per-request cap that differs by model version, so check yours. The keep clause is split by source. What comes from image one, what comes from image two, and which one wins on light.

Selfies deserve their own treatment because faces are where keep clauses fail most visibly, and ai selfie prompts covers the tells.

The Keep Clause, Expanded

A keep clause is a list of nouns that must survive the edit, and I'd write it out even when it feels redundant, because the model's default on an ambiguous edit is to regenerate the whole picture with your change included, which loses the face, moves the horizon, and swaps the background for a nicer one, all in a single pass that looks at first glance like it worked. The nouns that drift most, in the order I'd list them, are face and expression, pose including hands, framing and crop, light direction, background, and then anything specific to the photo such as a logo, a sign, or a particular object you can't lose.

The clause has a second use. Diagnosis.

If you wrote a keep list and the face still changed, the change you asked for was too big to be an edit and the model treated it as a generation. Split it. Change the jacket in one pass, then the background in another. Two clean edits beat one ambitious prompt, and confirming that costs you two renders.

One thing worth knowing about the Google models specifically is that the docs say all generated images include a SynthID watermark. It's invisible and it's there whether the picture was generated or edited, which matters if the edited photo is going somewhere that checks.

Fixes, By Symptom

Sorted by what you see in the output rather than by what you wrote, because the output is what you have.

The "edit ignored" row is the one I'd explain first, because it isn't obvious. Some edits are too large to be edits. Turning a daylight street into a night street isn't a change to the photo, it's a request for a different photo that happens to share a layout, and the model either refuses quietly or generates something new. The honest fix is to shrink the ask.

Google's guidance for photorealism is to describe the scene in rich detail, and that the more specific you are, the more control you have. That's true for generation and it's dangerous for edits, where specificity should be spent on the one change and the keep list, not on redescribing a scene the model can already see. The photographic vocabulary for the generation side, grain, stock, lens behaviour, is in ai photo prompt.

Questions From The Results Page

What is a good prompt for an AI image? One with all four slots filled. Subject with a particular, setting with a light direction, a look named as a medium or lens rather than a quality word, and a ratio. The market stall prompt at the top is the shape.

How do I do an AI prompt for my own photo? Reference, one change, keep clause. Start from Google's inpainting sentence and swap in your change and your keep list. If you want two things changed, run two prompts.

Are there fun AI prompts for photos? Yes, and the good ones are the ones that produce something you'd send to someone. I've collected those separately in fun ai prompts, with the image ones in their own section.

What are the trending photo prompts? Whatever was on the reels this week. It'll be different next week. The trends are restyles, and every restyle is the style transfer prompt above with a different style word, so once you have the shape you can do this week's trend and next week's without waiting for someone to post the string.

Where can I find free AI picture prompts? The banks on this results page are free and large, and they're useful as vocabulary. Read before you paste, though. The biggest one sorts its prompts into categories that include lingerie, and the models it names are specific enough that a string tuned for one will drift on another. What doesn't drift is the structure.

What I'd Try First

Take a photo you already have and write the inpainting sentence for it. One change, then the keep list, five nouns minimum.

Run it. Then run the same change with no keep clause and put the two results side by side. The difference between them is the whole argument of this post, and it's a difference you'll be able to see from across the room.