Face Swap Artifact Guide: Eyes, Hairline, Jaw, Mouth, and Skin

A face swap can look acceptable at thumbnail size and fail the moment you zoom in. One eye points in a different direction. The hairline has a soft halo. The jaw looks pasted over a collar. Teeth repeat. Skin color changes inside an oval around the face.

These are not all the same failure. Each region carries different information, so each artifact suggests a different next step.

The most efficient rule is simple: identify the region, name the defect, and change one input variable at a time. Repeating the same generation without a diagnosis may produce a different random variation, but it rarely fixes a structural problem.

First Separate Four Failure Classes

Before looking at individual regions, decide which class best describes the problem.

1. Target-image failure

The uploaded scene lacks usable information. Typical causes are a small face, motion blur, extreme angle, hard shadow, overlap, or severe compression.

2. Reference-image failure

The face reference does not provide clear identity detail or does not match the target pose and expression. A profile target with a front-facing reference is a common mismatch.

3. Model limitation

Both inputs are reasonable, but the scene is difficult: reflective glasses, a hand across the mouth, dramatic colored light, an extreme expression, or several overlapping faces. A different input may help, but no setting can guarantee a perfect reconstruction.

4. Display or export failure

The generated file is acceptable, but a later screenshot, messaging app, resize, or low-quality JPEG export introduces blur, ringing, or color changes. Always inspect the original downloaded result before blaming the generation.

The Regional Diagnostic Table

Use the table to choose the first experiment, not as a guarantee. Several causes can exist at the same time.

Eye Artifacts

Eyes are often the first place viewers notice that an image is synthetic. Humans are highly sensitive to gaze direction and symmetry.

What to inspect

At 100% zoom, compare:

• iris direction;
• eyelid shape;
• distance between eye and eyebrow;
• reflections in both eyes;
• the boundary of glasses; and
• whether one eye is sharper than the other.

Why eyes fail

A target captured mid-blink may provide only one complete eye. A strong three-quarter angle makes the far eye much smaller. Glasses add reflections that the model can mistake for iris or eyelid boundaries. Motion blur removes the thin contrast edges needed to reconstruct the eye.

Best fixes

1. Use a target frame where both eyes are open and interpretable.
2. Match reference gaze and head angle more closely.
3. Prefer non-reflective glasses or a frame without glasses when the accessory is not essential.
4. Avoid artificial eye-enhancement filters before upload.

Do not copy one generated eye over the other. Perfect mirroring usually looks unnatural because real perspective, light, and eyelid shape differ between sides.

Hairline and Forehead Halos

A visible oval around the forehead usually means the model could not blend identity features with hair and lighting.

Common causes

• dark hair against a bright background;
• hair falling across the forehead;
• a hat or hard shadow;
• a crop that cuts too close to the scalp;
• a reference with a very different forehead or hairline; and
• aggressive portrait blur around hair strands.

Best fixes

Use a wider crop that includes the complete head and some background. If possible, choose a target where the hairline is visible and not crossed by many loose strands. Match the reference angle; a straight-on reference does not explain how the forehead should recede in a strong side view.

Avoid simply blurring the halo after generation. Blur can hide a hard line at thumbnail size while creating a conspicuous soft patch at full size.

Jaw, Chin, and Ear Problems

The lower face is where identity, head shape, neck, clothing, and background meet. It is also where group photos create overlap.

Typical warning signs

• a second jaw line beneath the first;
• a chin that floats above the neck;
• skin blending into a collar;
• an ear that disappears or appears twice;
• a beard ending in a straight edge; and
• the neighboring person’s cheek merging with the target.

Likely causes

The target may be turned too far, the chin may be cropped, or an object may cross the face boundary. A large difference in apparent face width between target and reference can also stress the blend.

Best fixes

Choose a frame with a clean jaw silhouette. Include the neck and shoulders in the crop. If another person overlaps the jaw, use a different photo rather than trying to paint over the overlap before upload.

For beards, use a reference and target with compatible facial-hair coverage when possible. The goal is not identical styling; it is avoiding a situation where the model must decide whether a dark edge belongs to identity, shadow, or hair.

Mouth and Teeth Artifacts

Open mouths contain complex, high-contrast details: lips, teeth, tongue, shadow, and sometimes motion blur. A person captured while speaking can have an asymmetric mouth that is difficult to map from a neutral reference.

What failure looks like

• too many or repeated teeth;
• teeth blending into the lip;
• one side of the mouth stretching farther than the other;
• a dark interior with no boundary; and
• reference facial hair covering the target mouth.

Best fixes

Match expression before matching every other property. If the target is laughing, use a permitted reference with visible teeth and a similar mouth opening. If that is not available, choose a calmer target frame.

Do not sharpen generated teeth aggressively. Sharpening increases local contrast but does not correct incorrect tooth count or lip geometry.

Skin Color and Texture Mismatch

A face-shaped patch with different color or texture is usually a lighting and processing problem rather than an identity problem.

Causes to check

• one side of the target is under colored light;
• the reference has flash while the target has soft light;
• the target already has a strong beauty filter;
• screenshots or social platforms removed color information;
• the face is underexposed and noisy; or
• the scene contains reflected light from clothing or walls.

Best fixes

Start with the original image. Apply only mild global exposure and white-balance corrections. Do not smooth only the face before upload; local smoothing can make the target skin texture inconsistent with the neck and background.

When evaluating the output, compare forehead, cheeks, neck, and ears. A face that matches itself but not the neck still contains an obvious seam.

Glasses, Earrings, Microphones, and Hands

Thin objects are structurally difficult because they cross identity and scene boundaries.

Glasses should remain part of the target scene, but the model may redraw their frame while changing the face. Earrings sit near the ear boundary. Microphones and hands may cover the mouth or jaw.

Use this decision rule:

• If the object is optional, choose a frame without it.
• If the object defines the scene, choose a frame where it covers as little of the face as possible.
• If the object is transparent or reflective, inspect both its shape and the content behind it.
• If the object is already blurred, do not expect the model to recover a clean edge.

Change One Variable at a Time

Keep a short experiment log. It can be as simple as this:

This method prevents a common mistake: changing crop, exposure, reference, and file format at once, then having no idea which change mattered.

When a Same-Input Retry Is Reasonable

A same-input retry can be reasonable when:

• the first result has a small stochastic detail problem;
• the target and reference both score well;
• no face is severely blocked;
• the intended face was correctly selected; and
• the defect is not repeated across several attempts.

Change the input instead when:

• the wrong face is selected;
• the same boundary fails repeatedly;
• one facial feature is absent from the source;
• the target is too small or blurred;
• the reference angle is clearly incompatible; or
• a hand, microphone, hair, or another person covers the boundary.

Retrying cannot reveal information that the source image does not contain.

Do Not Repair a Harmful or Misleading Output Into Credibility

Post-processing can be appropriate for harmless creative cleanup, but it can also make an AI image easier to mistake for a real photograph.

Before polishing, ask:

1. Do I have the right and consent needed to use the depicted person?
2. Could the result be interpreted as evidence of a real event?
3. Am I removing visible AI artifacts without adding a clear AI-edit label?
4. Does the image involve a sensitive setting such as politics, crime, health, employment, or private conduct?

If realism increases the risk of deception, do not publish the image as an unlabeled photograph. Review our AI Disclosure and Content Policy.

Acceptance Checklist Before Downloading or Sharing

Inspect the original generated file and confirm:

• both eyes point plausibly and eyelids are complete;
• hairline has no obvious oval, halo, or cutout edge;
• ears and jaw appear once and connect naturally to the neck;
• lips and teeth have plausible geometry;
• face color and texture match ears and neck;
• glasses, jewelry, hands, and microphones are structurally intact;
• nearby faces and background objects were not changed unexpectedly;
• the image has an appropriate AI-edit or parody label for its context; and
• the output complies with the rights and safety rules that apply to the source image.

A technically cleaner face is not automatically a safe or lawful image. Quality review and context review are separate gates.

Final Troubleshooting Order

When a result fails, use this order:

1. Inspect the original downloaded output, not a screenshot.
2. Name the failed region and defect.
3. Check whether the target contains enough information in that region.
4. Compare target and reference pose or expression.
5. Remove ambiguity with a better crop.
6. Change one variable and generate once.
7. Stop if the scene remains structurally unsuitable.

Start with the photo quality scorecard if you are unsure whether the source is worth another attempt. For group inputs, follow the target selection and cropping workflow.

When you have a suitable target and reference, use the faceswap tool and review the result at full size before sharing it.