How to keep a face and fine details consistent
Practical ways to keep a face, a product, and small details recognizable in an AI image-to-image edit. What to say, what to crop, and what this editor will not do.
Consistency means the same picture, not a new person
Keeping a face consistent means the result should still be the person or character in the photo you uploaded. It does not mean borrowing a face from somewhere else, and it does not mean asking the model to wear someone else's identity. This site is an image-to-image editor. It restyles and relights a frame. It is not a face-swap tool, and these steps will not turn it into one. Upload only photos you have the right to edit.
The model redraws the frame from your image plus your words. Anything you do not protect in the prompt is allowed to drift: an earring, a logo edge, the exact shape of a smile. Anything you overload with competing style words is also allowed to drift, because the sentence no longer has a clear subject. The work is to say the few things that must stay, and to stop talking after that. The free editor shows the source next to the result so you can do that comparison on purpose.
Give the face enough pixels
A likeness lives in a small set of landmarks: hairline, eye spacing, nose, mouth, glasses, jaw. If those landmarks are tiny in the upload, the redraw has nothing solid to follow. On the free path the long edge is 512 pixels. A face that occupies a small corner of a group photo becomes a suggestion, and suggestions get replaced by a generic face in the chosen style. Crop until the head and shoulders, or the product, fill most of the frame.
Leave a margin. Cutting through the hair or the top of the head makes the model invent a new outline. Soft, even light beats dramatic shadow when the goal is recognition. You can add drama in a later run, after a calmer edit has proved the face can hold. Blurry sources, heavy beauty filters, and sunglasses that hide the eyes all remove landmarks. Pick a clearer file if you have one. No prompt repairs a source that never contained the detail.
Name the landmarks, then stop
Write the keep as concrete nouns. “Same face shape, same glasses, same haircut, same mole on the left cheek” is something you can check. “Same vibe” and “make them look like themselves” are not. The model does not have a private notion of the person beyond the pixels and your sentence. If a detail matters, put it in the sentence. If it does not matter, leave it out so it does not compete.
Change one thing. A new medium, or a new light, or a new background. Not all three. “Watercolor, same person, same pose, plain paper background” is a controlled test. “Watercolor, new clothes, rainy Tokyo, cinematic, younger” is several edits, and the face is the part that absorbs the confusion. After the face holds, a second run can change the room. The prompt guide is a longer version of this rule. The anime guide and the paint guide apply it to those media.
Use Gentle when the likeness is the point
Gentle stays closer to the photo. That is the setting for a portrait, a pet, or a product whose shape is the product. Balanced is the default for most edits and is fair when the source is large and clear and you want the style to read immediately. Bold lets the prompt restyle more. Use it for a scene where you would accept a related drawing, not for a face you must recognize tomorrow.
If Gentle looks too close to the original, add a more specific medium phrase and run Gentle again before you move to Balanced. Specific words do more than a stronger change amount. “Flat cel color, thin outlines, same glasses” moves the surface while the landmarks stay put. Turning the dial to Bold and hoping is how you spend a free run on a stranger. You get two standard runs per UTC day without an account. Treat the first as a test of the keep, not as the final poster.
Small details need a decision
Hands, text, jewelry, and distant logos are unreliable in a redraw. Decide before you upload whether each one is essential. If a hand gesture is the picture, crop so the hand is large and mention it: “same hand position, same mug.” If the hand is incidental, crop it out or accept a simpler shape. If a label must stay legible, this is the wrong tool. The model repaints regions. It does not set type. Say “simple unlabeled bottle, same shape” rather than “keep the exact brand sentence.”
Patterns on cloth often melt into a texture. If the pattern is the point, say its structure: “same striped shirt, horizontal navy stripes.” If it is not, do not spend words on it. Pets are the same problem as people. Name the markings that make that animal that animal: “same dog, same white muzzle, same dark ear.” A style preset will still simplify fur. Your job is to notice whether the markings survived, not whether the fur looks expensive.
Compare, then change one variable
Put the result beside the source and check a short list: hairline, eyes, glasses or their absence, mouth, and one clothing fact. If two of those moved, the edit failed even if the style looks finished. Do not “fix” it by adding more style words. Remove words, crop tighter, or drop from Balanced to Gentle. Change a single variable so you know what helped.
The higher-fidelity Pro model, listed on the pricing page, is for when the standard result is close and you want more of the source’s structure held. It does not grant a different kind of edit, and it is not a way to replace a face. Free runs stay on the standard model. The AI image editor page uses the same controls. If you are still deciding whether a photo should be the start at all, read image-to-image versus text-to-image before you spend the runs trying to force a composition the upload does not contain.
Questions
Does Gentle guarantee the same face?
No setting can promise a perfect likeness. Gentle stays closer to the photo, and a large, well-lit face with a short keep-list is the setup that fails least. A tiny face, a Bold change, or a prompt that asks for a different person will drift.
Can I keep the body and use a different face?
No. That is a face swap, and this editor is not for that. The photo you upload is the subject. Describe a medium, a light, or a setting, and say which features of that photo must stay.
What if glasses or a logo disappear?
Name them in the prompt and make them large in the crop. “Same round glasses” is checkable. If the logo is small type, the redraw will not preserve the letters. Simplify the label or keep that version in a normal editor.