Why Your Face Swap Looks Fake, and 7 Fixes That Take a Minute

You uploaded two photos and got back something wrong. The eyes sit in roughly the right place, but the face still reads as pasted on. That is rarely the model failing at random. It is usually one of seven input problems, and each has a fix that takes about a minute.

Find your fault in this table

Start here. Match what you see in the output to the row below, then jump to the section that explains it.

What you see	Likely cause	Fix
Face looks stuck on, with a visible line along the jaw	Light hits from a different side in each photo	Pick a source photo lit from the same side, then set Blend Strength to 0.85
Skin tone jumps at the hairline or neck	One photo is warm, the other is cool	Use a source shot in similar light and turn on Face Enhancement
Result is soft, smudged or low detail	The face is too small inside your source photo	Crop tighter so the face fills more of the frame
Some video frames still show the original face	The head turned away or blurred, so no face was detected	Trim the clip to the seconds where the face stays front on
The wrong person got swapped in a group shot	Face Index 0 targets the largest detected face	Change Face Index, or switch on Swap All Faces
Eyes look flat, dead or slightly off centre	Glasses reflect light over the eye landmarks	Use a source photo without glasses
A strand of hair sits under the new face, not over it	Hair crossed the cheek in the target image	Choose a target where the hair is tucked back
The swap is clean but there is a logo in the corner	Free results are watermarked	Nothing is broken. That mark comes off on a paid plan.

Light direction is the tell everyone misses

Your brain reads light before it reads faces. When the source face is lit from the left and the target is lit from above, the join looks wrong even when the shape is perfect. The blend is fine. The lighting is not.

Here is a two second test. Look at the shadow under the nose in your source photo, then in your target. If one falls to the left and the other falls to the right, you have found your problem.

Flat, even light is the safest source you can pick. An overcast day, a window on a grey afternoon, or an indoor shot with no single bright lamp all work well. Hard sunlight and phone flash are the two worst source photos, because both carve strong shadows the AI has to fight.

The 30 degree head angle rule

Face swap models work from facial landmarks. Those are the points around the eyes, nose and mouth. Turn the head far enough and half of them disappear behind the cheek.

As a rule of thumb, stay inside about 30 degrees of straight on. You do not need a protractor. Just check whether you can see both eyebrows clearly. If the far eyebrow is hidden or clipped by the nose, the head has turned too far.

Photos and video fail differently here. A still photo gives you one sliding, smeared face. A clip behaves better than people expect: on the video face swap tool, frames where no face is detected pass through untouched. That is why a swap can look great for three seconds and then snap back to the original person mid turn. Nothing crashed. Those frames simply had no face to work with.

Anything covering the face costs you detail

Glasses

Lenses bounce light straight back at the camera. That reflection sits exactly where the eye landmarks are, so the model guesses. Guessed eyes are why a swap can look almost right yet feel lifeless. Use a source photo with no glasses, even if you wear them daily. Thin frames are usually fine. Reflections and tinted lenses are not.

Hair

Hair crossing a cheek causes a specific, obvious failure. The swap replaces the face region, so the strand ends up buried under the new face instead of lying on top of it. Pick a source and a target where the hair is pushed back off the face. A ponytail shot beats a windy beach shot every time.

Hands, masks and microphones

Same rule, same reason. A hand on the chin, a mic in front of the mouth or a mask over the nose removes landmarks the model needs. Crop or choose a different frame. No setting recovers a covered face.

Fast scenes drop frames to motion blur

Motion blur smears the eyes and mouth into the background. Detection then fails, and the original face stays. In a fast cut clip you can lose a run of frames this way, which looks like flickering between two people.

Two things help. Trim to the calmest few seconds before you upload, and raise the frame cap so the tool actually processes the whole clip. The video tool lets you pick how many frames to run.

60 frames60

150 frames150

300 frames300

600 frames600

1440 frames1440

The five Max Frames settings listed on the video face swap page. Each bar shows that cap as a share of the 1440 frame maximum.

Frame count adds up fast. A 10 second clip at 24 frames per second is 240 frames, and the model runs once on every one. That is also why a video takes about two minutes while a photo takes seconds.

Small source photos cannot be rescued

This is the fix people skip, and it is the one that matters most. The photo face swap tool asks for a source image of at least 512 x 512 pixels. The face inside it should be at least 200 x 200. A face that occupies 60 pixels of a group photo has almost no detail to copy.

Crop before you upload. Cut the source photo down so the face fills roughly a third of the frame or more. You are not adding detail, but you are stopping the tool from working with a tiny region.

Bigger is not automatically better, though. Do not feed the video tool a 4K file. The face is analysed at a fixed internal size whatever you upload, so 4K buys no accuracy and eats your upload limit. Stay at 1080p or below. Keep still images under about 10 MB.

Two faces in the frame, one wrong swap

Group photos have a rule you can predict. Face Index 0 means the largest detected face, so the person nearest the camera wins by default. Your friend at the back gets ignored.

You have two ways out. Set Face Index to target a different face, or turn on Swap All Faces and replace everyone at once. Both are on the upload panel, not buried in an advanced menu.

One limit is worth knowing before you plan a group edit. You supply one source face per job. Every face that gets swapped becomes that same identity. Different faces for different people means running the tool once per identity.

Three settings worth changing before you re-run

Fix the photo first. These settings clean up what is left.

Blend Strength

The default is 1.0, a full replacement. That is cleanest on a still, calm face. Drop it to 0.85 when you can see a hard edge, and the join softens without losing the effect. Going below about 0.7 starts ghosting the original face through, which reads as a double exposure.

Face Enhancement

This runs a sharpening and restoration pass over the swapped face. It costs a few extra seconds and fixes most soft, plasticky results. Leave it on for anything you plan to post.

Output Quality

High adds another sharpening pass. On video, High applies enhancement to every frame, so processing takes longer. Use Standard while you test and High for the version you keep. Running a test at Standard first also saves you tokens, which matters more once you are on a paid plan with a monthly token balance.

Run this check, then swap again

Go through this before the upload, not after the disappointing result.

Work in this order and you will spend less time re-running jobs. Fix the source photo, then re-run at Blend Strength 0.85 with Face Enhancement on. Still off? Change the target, not the settings, because a target with a turned or shadowed head will resist every slider you touch.

Test on a still image before you commit to a clip. A photo swap finishes in seconds. It tells you straight away whether your source face is good enough. Once it looks right, move the same photo to the GIF tool or the video tool. Want a quick sanity check on the photo itself? The AI age guesser is a decent proxy. A photo it cannot read is a photo the swap will struggle with too. For the rest of the image tools, the how-to guides cover cropping, upscaling and background removal.