For a more controlled image edit, state the single change you want, then explicitly name what must stay the same. Describe the person’s likeness, pose and expression, plus any scene details that matter; tell the tool to change only the target. This makes the request clearer, but it cannot guarantee that every detail will be preserved.
What to include in an image-editing prompt
Think of the prompt as a compact specification: one requested change, followed by the visual constraints that define a successful result. OpenAI’s image-prompting guide recommends naming the target of an edit and listing what should remain unchanged.
As an Amazon Associate I earn from qualifying purchases.
- The requested change: Identify the exact operation, such as replacing a jacket, removing a cup, or changing the color of a wall.
- Identity and appearance: For a person, specify likeness, facial features, skin tone, body shape, hairstyle and proportions when those details matter.
- Pose and expression: State the person’s posture, gesture, gaze and expression. If they are holding or interacting with something, describe that relationship.
- Composition and surroundings: Name the framing, camera angle, background, lighting, layout, labels or nearby objects that should stay fixed.
- Boundaries: Say “change only” the target and mention unwanted additions, such as extra text, accessories, logos or watermarks.
Choose constraints that are relevant to the edit rather than listing every possible detail. For a local change, mention the target and the nearby elements that must not move or change.
A prompt pattern for a targeted edit
Adapt this template to the image and the editing tool:
#1 Best Overall
Edit the supplied image to [one specific change]. Preserve [identity and/or subject attributes], [pose and expression], and [composition and scene details]. Change only [target element]. Keep [lighting, shadows and color] consistent with the source. Do not add [specific unwanted elements].
For example, to remove a distracting object, specify which object to remove and what should remain around it: “Remove the red cup from the table. Preserve the person, their pose and expression, the table, the background and the original lighting. Change only the cup and the area it covers; do not add new objects or text.”
The template is a practical way to apply the guidance, not a quoted formula or a guarantee of the result. OpenAI’s guide gives examples of changing clothing while preserving a person’s face, features, skin tone, body shape, pose, identity, expression, hairstyle and proportions, and of removing an object while keeping surrounding details intact.
Recommended Free Tools
How to combine multiple reference images
When an edit uses more than one image, give each image a clear role. Say which is the base scene and which supplies the person, clothing, object or style. Then specify what to transfer and what to preserve from the base image.
Image 1 is the base scene; Image 2 is the clothing reference. Apply the clothing from Image 2 to the person in Image 1. Preserve the person’s identity, pose, framing and background, and match the original lighting.
Explicit roles help distinguish the source of each requested element; simply attaching several images may leave their intended uses unclear. OpenAI’s GPT Image Generation Models Prompting Guide also discusses structuring image prompts, but the precise controls available depend on the tool and model.
How to iterate when details drift
Inspect each result before making another request. If something changed that should have stayed fixed, identify that specific drift and ask for one correction at a time. Restate the important constraints—such as likeness, pose or framing—rather than adding several unrelated changes to the next edit.
- Review the output against the original and identify the most important mismatch.
- Ask for one targeted correction, naming both the item to fix and the details to preserve.
- Use the edited result as the input for the next edit when the tool supports it.
- Inspect again before moving on to another correction.
This keeps the next instruction focused and makes it easier to see whether the specific issue improved. Consistency is still something to check: the ICCV 2025 paper Edicho: Consistent Image Editing in the Wild addresses image-editing consistency as a research problem. The sources cited here do not establish a general success rate for any prompt wording.
Best Value
How to judge whether an edit preserved what matters
Compare the output with the original using the constraints you named. For a person, check likeness, facial features, skin tone, body shape, hair, proportions, pose and expression. For the scene, check framing, camera angle, background, lighting, layout, labels and nearby objects. Also check for unwanted additions or changes beyond the requested target.
These checks are useful when comparing tools, too. Consider how well each handles identity-sensitive edits, local changes, reference-image roles, unwanted changes and correction across iterations. OpenAI’s guide names instruction following, identity and product preservation, text accuracy, unwanted changes and transparency as evaluation areas; that list is a checklist, not independent evidence that one tool will perform best.
Tool-specific requirements are not universal prompt rules
Input limits and supported formats depend on the service. For example, the OpenAI image edit API reference lists PNG, WebP and JPEG inputs under 50 MB for the GPT Image models it covers. Those are API-specific requirements, so check the current reference for the model and endpoint you use rather than treating them as rules for all image editors.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
