Photo Editing AI Prompts: 2026 Guide

Author: AvocAIdo · Published: September 7, 2026

AI photo editing prompts are text instructions that tell an AI how to modify an existing image while preserving selected details. Effective prompts identify the target, requested change, desired appearance, protected elements and prohibited changes. This guide explains that structure, provides copy-ready examples and shows how to troubleshoot common failures across popular AI editing workflows. If you are new to the process, our step-by-step guide to editing photos with AI covers the foundations before you dive into prompt technique.

AI photo editing prompts in short

• An editing prompt changes an uploaded image, while an image generation prompt creates a new composition from text.
• A reliable prompt has 5 parts: action, target, desired result, keep clause and negative constraint.
• "Change only the jacket; keep everything else the same" is more reliable than "improve this fashion photo."
• For multiple references, identify each image by number and state exactly which elements should move or influence the result.
• Use 1:1 for square posts, 16:9 for web banners and 9:16 for Stories or Reels.
• Complex edits are easier to control through 2 or more focused passes rather than one overloaded instruction.
• A prompt library provides starting points, but identity preservation and brand consistency require image-specific constraints.

Checklist for writing a better editing prompt

• Upload the highest-quality version of the original image.
• Name the exact object, person or region to edit.
• Describe one primary change per editing pass.
• Specify the intended visual result and level of realism.
• State which facial, product and background details must remain unchanged.
• Add negative instructions for likely failure modes.
• Match the new lighting direction to the original photograph.
• Define the required aspect ratio before expanding the composition.
• Inspect faces, hands, text, shadows, reflections and edges at full size.
• Repeat the preservation clause during every refinement pass.

What can photo editing AI prompts change in 2026?

Prompt-based editors can replace backgrounds, adjust lighting, restyle clothing, remove distracting objects, expand a composition, repair an old photograph or create a new visual treatment from an uploaded image. The defining feature is that the user describes the desired change in natural language instead of performing every adjustment manually. To understand how AI photo editing actually works at a technical level, see our dedicated explainer.

This approach is especially useful when the requested outcome can be described clearly. Examples include changing a grey studio background to warm beige, converting midday light into golden-hour light or replacing a jacket while preserving the model's face and pose.

Common applications include:

• Portrait and professional headshot refinement
• Fashion and lifestyle campaign variations
• Background replacement and scene expansion
• Product and food photography cleanup
• Interior staging and architectural visualization
• Vintage photograph restoration and colorization
• Cinematic, fantasy, anime or editorial restyling
• Social media resizing and aspect-ratio expansion

The output specification can also be part of the instruction. For example, an editor may be asked to expand a landscape image to 16:9 or prepare a vertical 9:16 composition without cropping the main subject. AvocAIdo also presents 1:1, 16:9 and 9:16 as common formats for Instagram, web banners and Stories or Reels respectively. (zdroj: AvocAIdo)

Prompt editing is not the same as applying a preset. A preset generally makes predetermined tonal changes across an image. A prompt can express relationships and constraints, such as matching the light on a newly generated background to the existing light on a woman's face.

How do editing prompts differ from generation prompts?

An image generation prompt asks a model to build an image from scratch. A prompt for AI photo editing starts with an existing photograph and requests a controlled modification. This distinction affects how the instruction should be written.

A generation prompt might say:

Photorealistic studio portrait of a woman in a dark green blazer, soft side lighting, neutral grey background, editorial fashion photography.

An editing prompt should instead say:

Change only the woman's black blazer to dark green. Preserve her face, hairstyle, expression, pose, hands, body proportions, studio lighting and neutral grey background. Keep the photograph realistic. Do not add accessories or alter skin texture.

The second version contains invariants, meaning details that must not change. Aireiter distinguishes editing from generation in the same way: an editing prompt modifies a photo supplied by the user, rather than building the complete image from text.

OpenAI's current image prompting guidance recommends the pattern "change only X" followed by "keep everything else the same." It also recommends repeating the preserve list during later iterations because successive edits can introduce drift.

This distinction matters for stylistic requests too. "Create a vintage portrait" gives the model broad creative freedom. "Apply a restrained 1970s film treatment while preserving the face, clothing, pose and composition" confines the change to color, grain and tonal character.

Use generation language when the entire scene is open to reinterpretation. Use editing language when the original identity, geometry, product design or composition must survive. For a deeper look at the multimodal AI models like Gemini and GPT Image that power these workflows, see our creative professionals' guide.

What is the anatomy of a strong AI editing prompt?

Several current guides describe closely related formulas. Picsart uses 4 components: action, target, desired look and what to preserve. Aireiter adds a fifth component, the negative constraint. ByteVerse condenses the principle to target, change and constraint.

For professional work, the 5-part structure offers the clearest control:

1. Action

Begin with a direct editing verb:

• Remove
• Replace
• Recolor
• Restore
• Expand
• Relight
• Retouch
• Recompose

In a selection-based tool such as Photoshop Generative Fill, the selected region already provides context. Adobe states that the feature uses a selection and prompt, then places the result on a Generative Layer with a mask for nondestructive editing. Adobe also documents that an empty prompt can let Photoshop fill a selection from surrounding pixels.

2. Target

Identify the exact subject or region. "Change the clothing" is vague. "Change only the navy jacket worn by the foreground subject" is spatially specific.

Useful identifiers include:

• Foreground or background
• Left, right or center
• Closest to the camera
• Behind the product
• Inside the selected region
• The second reference image
• The visible sleeve, label or tabletop

3. Desired result

Describe what the finished edit should look like. Include the material, color, lighting, photographic style or mood that directly affects the result.

For example:

Replace the white tabletop with pale honed limestone, subtle natural texture, matte finish and realistic contact shadows.

Modifiers such as "ultra" or "cinematic" are not substitutes for concrete direction. "Photorealistic, soft window light from camera left, shallow depth of field" gives the editor more useful information than "ultra stunning lifestyle image."

4. Keep clause

The keep clause is the control center of an editing prompt. Name the features that must remain unchanged:

Preserve the face, identity, expression, hairstyle, pose, hands, body proportions, camera angle and background.

For products, preserve shape, dimensions, logo placement, packaging colors and label hierarchy. For interiors, preserve walls, windows, doors, floor plan and camera perspective. Documenting these constraints in AI-ready brand style guide templates makes them reusable across every editing session.

5. Negative constraint

Finish with likely failure modes:

Do not beautify the face, smooth the skin, alter the fingers, add jewelry, modify the background or generate readable text.

Negative instructions should be specific rather than generic. "No errors" does not tell the model what an error would be. "No extra fingers, warped earrings or duplicated buttons" does.

A reusable master template is:

[Action] only [precise target] to [desired result]. Match [lighting, perspective, texture or style]. Preserve [protected details]. Keep everything else unchanged. Do not [specific unwanted changes]. Output as [ratio or background requirement].

How do you edit a photo with AI step by step?

A controlled workflow separates creative intent from technical correction. Use the following process whether the editor accepts a full photograph, a selected region or several reference images.

1. Define one primary objective. Decide whether the pass is for background replacement, clothing modification, restoration, relighting or another single outcome.
2. Identify protected details. List the elements that make the image usable: identity, expression, product geometry, label, composition, pose or architectural layout.
3. Select or describe the target precisely. If masks are available, select the relevant region. Otherwise use spatial language such as "the red jacket on the central subject."
4. Write the requested result. Specify color, material, setting, lighting direction, realism and intended use. Avoid stacking unrelated aesthetic terms.
5. Add a keep clause and negative constraint. State what remains identical and name the 2 or 3 most likely unwanted changes.
6. Generate an initial edit. Evaluate the requested change before judging small details. Confirm that the AI edited the correct target and respected the composition.
7. Refine one defect at a time. If the face changed, restore identity before adjusting color. If the new background works but the shadows do not, make shadow matching the next pass.
8. Inspect the full-resolution image. Examine eyes, teeth, fingers, hair edges, logos, text, reflections and object contact points. These regions commonly reveal inconsistencies.
9. Export for the intended channel. Use an appropriate ratio such as 1:1, 16:9 or 9:16, and keep an editable original before final delivery.

For old photos, ByteVerse recommends the sequence restore, optionally colorize, then upscale. Upscaling first can enlarge and sharpen existing defects. Its guide suggests a restrained 2x to 4x upscale rather than automatically choosing a maximum setting.

Which prompts preserve faces and identity?

Identity preservation requires more than adding "same person." Describe the stable visual features and prohibit beautification or structural changes. This is particularly important for professional portraits, documentary images and campaign photography.

Portrait cleanup

Refine exposure and white balance across this portrait. Preserve the person's exact facial identity, facial proportions, expression, eye shape, nose, lips, natural skin texture, hairstyle and age. Retain realistic pores and fine lines. Do not beautify, reshape the face, whiten the teeth or create plastic skin.

Professional headshot

Replace only the background with a clean warm-grey studio backdrop. Keep the subject's face, identity, hair, clothing, pose and body proportions unchanged. Match the existing light direction and preserve realistic hair edges. Do not modify facial features or add accessories.

Selfie to studio portrait

Convert the lighting and background of this selfie into a professional studio portrait. Use soft key light from camera left and a subtle neutral background. Preserve the exact face, expression, hairstyle and clothing. Correct lens distortion conservatively. Do not slim the face or smooth away natural skin texture.

Fashion clothing change

Change only the black jacket to a tailored burgundy wool blazer. Keep the woman's face, identity, pose, hands, body proportions, shirt, trousers and studio background identical. Preserve the original folds and light direction. Do not add jewelry, pockets, buttons or text.

For subsequent passes, repeat the full identity clause rather than saying "fix the face." OpenAI specifically recommends repeating invariants on every iteration to reduce drift.

Consent should be part of the workflow before editing a recognizable person. Obtain appropriate permission for commercial likeness use, take additional care with images of minors and avoid presenting synthetic changes as documentary truth. Do not use a real person's likeness to imply an endorsement, event or action that did not occur.

Which prompts work for portraits, products and brands?

Copy-paste libraries are most useful when treated as editable frameworks. Replace every bracketed placeholder with image-specific details. To make your AI-edited photos look professional and on-brand, pair these templates with documented visual standards.

Lifestyle portrait

Replace only the indoor background with a bright modern café during late morning. Match the original camera height, perspective and light direction. Preserve the subject's face, identity, clothing, pose and hands. Add realistic contact shadows and a restrained depth of field. Do not alter body proportions or add readable signs.

Editorial fashion photograph

Apply a high-end editorial fashion treatment through controlled contrast, muted cool shadows and clean skin tones. Preserve the model's identity, garment design, stitching, silhouette, pose and studio composition. Keep fabric texture realistic. Do not reshape the body, replace accessories or generate typography.

Vintage film treatment

Restyle the photograph with a subtle 1970s color film appearance: warm highlights, gently faded blacks, restrained grain and slightly muted saturation. Preserve the full composition, face, clothing and original objects. Do not add scratches, light leaks, borders or artificial date stamps.

Old photograph restoration

Restore this damaged photograph by reducing scratches, dust, stains and faded contrast. Reconstruct only visibly damaged regions using surrounding detail. Preserve every person's identity, age, expression, clothing and the original composition. Do not modernize hairstyles, invent objects or over-sharpen faces.

Product background removal

Remove the existing background and isolate the complete product on transparency. Preserve the exact product shape, proportions, packaging colors, logo placement and edge detail. Retain a clean natural edge without halos. Do not redesign the label or generate new text.

Marketplace product image

Place the unchanged product on a pure, uncluttered studio background with soft natural contact shadow. Preserve the packaging geometry, cap, material texture, colors and label placement. Keep the product centered and fully visible. Do not add props, claims, badges or altered text.

Food photography

Improve only the lighting and color balance of this food photograph. Create soft side light, appetizing but natural color and realistic texture. Preserve the plating, ingredients, portion size, tableware and camera angle. Do not add garnish, steam or ingredients that were not present.

Interior photograph

Replace only the empty rug area with a low-pile beige wool rug that follows the room perspective. Preserve the floor plan, walls, windows, doors, furniture, lighting and camera position. Add correct contact shadows. Do not move furniture or alter architectural dimensions.

Fantasy or anime restyling

Reinterpret the uploaded portrait as a detailed fantasy illustration while preserving the recognizable face, expression, hairstyle, pose and framing. Use painted moonlight, deep blue shadows and restrained silver accents. Do not change age, body proportions or background layout.

A brand prompt should begin with documented visual standards. AvocAIdo's brand workflow recommends defining preferred color palettes, lighting, composition rules and mood descriptors before editing.

How should the same prompt change across AI platforms?

The core request can remain consistent across tools, but the amount of context changes. Conversational editors benefit from an explicit action and preservation list. Selection-based workflows already know which pixels are targeted, while reference-based generation needs a clear relationship between inputs. For an honest comparison of the best AI photo editing tools in 2026, see our dedicated review.

Workflow	Adapted instruction for the same jacket edit
ChatGPT with GPT Image	"Edit the uploaded photo. Change only the black jacket to a burgundy wool blazer. Preserve the exact face, identity, pose, hands, lighting and background. Keep everything else the same."
Google Gemini	"Using the uploaded photo as the source, replace only the black jacket with a burgundy wool blazer. Keep the person's face, expression, pose, hands and background unchanged. Do not modify any other element."
Adobe Firefly	"Burgundy tailored wool blazer, matching the original folds, perspective and studio lighting. Preserve the subject and surrounding composition."
Photoshop Generative Fill	After selecting the jacket region: "Burgundy tailored wool blazer, realistic wool texture, same folds and light direction."
Midjourney reference workflow	"Editorial fashion portrait, same subject pose and composition as the image reference, burgundy tailored wool blazer, matching studio light, realistic fabric texture."

The Photoshop version can be shorter because a precise selection supplies the target. Adobe's official documentation confirms that Generative Fill works from a selected area and can create the result on a masked Generative Layer. (zdroj: Adobe)

For GPT Image workflows, explicit invariants should remain in the prompt. OpenAI advises identifying exactly what may change and what must stay fixed. When several input images are used, its guidance recommends referring to each one by index and description. (zdroj: OpenAI)

For example:

Use Image 1 as the base photograph. Use only the jacket material and burgundy color from Image 2. Apply that material to the jacket in Image 1. Preserve the face, pose, hands, trousers, lighting, framing and background from Image 1.

A Google Gemini AI photo editing prompt can use the same structure. The important part is not a platform-specific magic phrase. It is the unambiguous relationship between the source image, reference image, editable target and protected details. Browse our full list of AI models for image generation to see which engines support reference-based editing.

How do reference images and localized edits improve control?

Reference images communicate visual information that is difficult to encode in adjectives. A reference may define a color palette, garment material, lighting treatment, product arrangement or full visual identity.

For multi-image editing, label the sources in order:

• Image 1: base photograph and composition
• Image 2: garment material reference
• Image 3: lighting reference
• Image 4: color-grade reference

Then describe how the inputs interact. OpenAI's image prompting guide recommends referencing each input by index and specifying which elements move or influence the final composition.

For localized edits, masks and selections reduce ambiguity. Select only the jacket when changing clothing, but include a small, useful boundary around its edges when the tool needs to reconstruct folds or hair overlap. Keep faces, logos and unrelated objects outside the editable region whenever possible.

Spatial language provides another control layer:

Change only the jacket worn by the central foreground subject. Do not edit the face above it, the hands overlapping the sleeves or the grey wall behind it.

Background replacement requires attention to light and contact. ByteVerse recommends choosing a new background lit from the same side as the subject and applying a shared global color grade. This helps the subject and generated environment appear to belong to one photograph.

For style consistency across several assets, use the same references, preserve list and lighting vocabulary. AvocAIdo combines reference photos with text instructions, and its Brand Style Intelligence analyzes attributes including color palette, lighting, mood and visual style. (zdroj: AvocAIdo) The feature is available across its plans and supports unlimited saved styles.

Why do AI photo editing prompts fail?

Most failures can be traced to an unclear target, too many simultaneous requests or missing preservation constraints. Use this prompt-first decision tree. Checking your tool against the 10 must-have features in AI photo editing tools can also reveal whether the platform itself is limiting your results.

The person's identity changed

Cause: The prompt described a general portrait transformation without locking facial details.

Fix:

Restore the exact identity from the original uploaded image. Match the original facial proportions, eyes, nose, lips, jawline, expression, age and natural skin texture. Change nothing else.

Repeat this preserve list during every next pass.

Fingers or hands look incorrect

Cause: Clothing or pose changes forced the model to reconstruct overlapping hands.

Fix: Return to the cleanest prior version. Edit the garment without requesting a new pose, and state:

Preserve the exact hand position, finger count, finger shape and sleeve overlap from the original.

Skin looks plastic

Cause: Terms such as "flawless," "perfect" or "beautified" encouraged excessive smoothing.

Fix:

Restore natural skin texture, pores, fine lines and subtle tonal variation. Preserve the original age and facial structure. Do not apply beauty-filter smoothing.

Text or a logo changed

Cause: The editable region included typography, or the model regenerated packaging.

Fix: Keep the label outside the mask when possible. Add:

Preserve the existing logo, typography, spelling, line breaks, label dimensions and placement exactly. Do not generate or rewrite text.

For important commercial copy, verify every character and consider adding final labels through a conventional design workflow.

Shadows do not match

Cause: The new object or background uses a different light direction.

Fix:

Match the original key light from camera left. Add a soft contact shadow directly beneath the product, with direction, density and softness consistent with the existing scene.

Reflections are inconsistent

Cause: Reflective surfaces were not included in the requested change.

Fix:

Update only the product reflection on the tabletop to match its current shape, color and position. Preserve all other reflections and the camera perspective.

The background changed unexpectedly

Cause: "Make this cinematic" gave the model permission to reinterpret the whole image.

Fix:

Apply the requested color treatment only. Lock the background, composition, objects, subject position and camera angle. Do not add, remove or move anything.

The result is overloaded

Cause: One prompt requested restoration, colorization, background replacement, wardrobe changes and upscaling together.

Fix: Divide the work into focused passes. For vintage material, complete restoration first, colorization second and upscaling last. A sequence of 3 controlled prompts is easier to inspect than one unrestricted transformation.

How do you request output-ready images?

Appearance is only one part of a usable deliverable. State the final composition, background behavior and intended channel at the end of the prompt.

Transparent background

Remove the complete background and retain transparency. Preserve the full product, including fine edge details and any intentional translucent material. No white fringe, colored halo, cropped edge or generated shadow.

Aspect-ratio expansion

Expand the composition to 16:9 by extending the background naturally on both sides. Keep the subject at the same scale and position. Do not crop, stretch or regenerate the face, body or clothing.

Vertical social version

Recompose to 9:16 by extending the scene above and below the subject. Preserve the original person, pose and camera perspective. Maintain sufficient uncluttered space around the subject without adding text.

Square product cover

Prepare a centered 1:1 product composition. Keep the complete product visible with balanced negative space. Preserve packaging geometry, colors, label placement and realistic contact shadow.

Print or marketplace preparation

A prompt may request clean edges, realistic texture, consistent lighting and an uncluttered background. However, technical delivery requirements should still be checked after generation. Inspect dimensions, color mode, transparency, typography and platform-specific marketplace rules rather than assuming that visual appearance confirms compliance.

For a commercial campaign, keep a record of the original source, references, approved edit and final disclosure decision. Review recognizable people, third-party branding and generated additions before publication.

AvocAIdo tip

If writing long model-specific syntax interrupts the creative process, describe the image change in plain language and retain a clear keep clause. AvocAIdo's ArtDirector AI converts plain-language briefs into professional prompts and is included across its AvocAIdo pricing plans. The platform then runs Grok Imagine, Seedream 5 Lite, NanoBanana 2 and GPT Image 2 in parallel, allowing the user to review 4 interpretations of the same direction. Reference-based Brand Style Intelligence can also analyze color, lighting, mood and visual style for repeatable campaign direction. Explore the workflow on the AvocAIdo homepage.

Frequently asked questions

What is the best prompt formula for editing a photo?

Use 5 parts: action, precise target, desired result, keep clause and negative constraint. For example, request one jacket change, preserve the face and background, then prohibit added accessories or altered skin. This structure gives the model both a task and clear boundaries.

Can I copy and paste a Gemini photo editing prompt?

Yes, but replace every placeholder and generic description before submitting it. A Google Gemini AI photo editing prompt copy paste template should identify the uploaded source, request 1 main change and list protected details such as identity, pose and background. Refine the result with another focused instruction rather than rewriting the entire concept.

How can I stop AI from changing a face?

Use "change only X" and repeat a detailed identity clause in every editing pass. Protect facial proportions, expression, age, skin texture and hairstyle, not merely "the same face." If drift appears after 2 edits, return to the cleanest earlier image rather than repeatedly repairing a degraded version.

Are free text-prompt photo editors sufficient?

A free AI photo editor text prompt workflow can be suitable for learning prompt structure and testing straightforward edits. Evaluate at least the full-size face, hands, edges, text, shadows and reflections before publishing. AvocAIdo offers a free trial with 4,000 credits and no credit card requirement.

Should I use negative prompts for photo editing?

Yes, when the unwanted outcome can be described precisely. Add 2 or 3 likely failure modes, such as "do not smooth skin, alter fingers or change the background." Avoid long lists of irrelevant negatives because they can distract from the main edit.

What aspect ratio should I request?

Use 1:1 for a square composition, 16:9 for a landscape web banner and 9:16 for vertical Stories or Reels. Ask the editor to expand the background instead of stretching or cropping the subject. Confirm that faces, products and important edge details remain intact after expansion.

How many changes should one editing prompt contain?

Start with 1 primary visual change per pass, especially when identity or product geometry matters. A complex restoration can be divided into 3 stages: restore damage, colorize if required, then upscale. Focused passes make it easier to identify which instruction caused an error.

Next step

Choose one photograph and write a 5-part brief covering the action, target, desired result, protected details and prohibited changes. You can then start with AvocAIdo to test a plain-language direction across its image workflow. The free trial includes 4,000 credits without requiring a credit card.

Zdroje

1. AvocAIdo, AI Photo Editing Explained
2. Adobe, Edit Images with Generative Fill
3. OpenAI, Image Generation Prompting Guide
4. AvocAIdo, Homepage
5. AvocAIdo, Pricing