How to Face Swap a Video: Every Step, Start to Finish

AskAI.free team
How to Face Swap a Video: Every Step, Start to Finish
The short version
  • You upload two files: one still photo of a face, and one video clip. The AI swaps the face on every frame and returns an MP4.
  • A free account handles clips up to 20 seconds and 5 MB. Pro and Max raise that to 60 seconds and 100 MB.
  • Most clips finish in about two minutes, because the model runs once per frame instead of once per video.
  • Swaps fail for a short list of reasons: side angles, glasses, hair over the face, motion blur, mismatched light, and a source photo that is too small.
  • Frames where no face is detected pass through untouched. A bad swap usually looks like flicker, not like a melted face.

You have a clip and a photo of a face. You want the result to look real, not broken. Most guides stop at "upload and click", which is the easy part. This one covers the settings, the waiting, and the specific reasons a swap goes wrong.

The two files you need before you start

Only two uploads matter. Everything else is a setting you can leave alone.

The source face must be a still image, either JPEG or PNG. It cannot be a video. The tool reads one face identity from that single picture, then reuses it on every frame. A video source is not allowed, and that is deliberate: one clean still beats an average of many blurry frames.

Your target clip can be MP4, MOV or AVI. Output always comes back as H.264 MP4, so it drops into YouTube, TikTok, Reels or any editor.

20 slongest clip on a free account
5 MBvideo file cap, free account
~2 mintypical processing time
1source face per job

Guests can try a single photo swap with no account. Video and GIF swaps need a free account, which keeps the queue moving for everyone.

Every step, in order

1. Pick the source photo carefully

This choice decides your result more than any slider does. Use a front-facing photo with even light. Both eyes should be visible. The face should fill at least a third of the picture, because that gives the model more detail to copy.

Skip sunglasses, heavy shadow and extreme angles. Shot it on an iPhone? HEIC files will not upload, so save as JPG first.

2. Trim the clip, then upload it

Cut the clip down to the moment you actually want. Black leader frames, title cards and slow fades all cost the same as real frames. Trimming those off a typical clip removes roughly 10 to 20 percent of the frame count, and that time saving is real.

Stay at 1080p or lower. A 4K file gives you no extra swap quality, because the face is analysed at a fixed internal size no matter what you feed it. All 4K does is eat your upload allowance.

3. Set the four controls

Four settings sit under the upload box on the video face swap tool:

  • Max Frames. Options are 60, 150, 300, 600 and 1440. The 300 setting is marked as the best balance.
  • Blend Strength. 1.0 is a full swap. 0.5 leaves the original face ghosted underneath at half weight.
  • Swap All Faces. Turn this on and every face in frame becomes your source identity.
  • Frame Quality. Standard or High. High runs face enhancement on every single frame, so it takes noticeably longer.

4. Submit, then walk away

Your job enters a queue and runs in the background. You do not need to keep the tab open. Progress shows as a percentage, and the finished file lands in My Files when it is done.

5. Download and check the audio

Free results carry a watermark. Adding that watermark means re-encoding the video, and the re-encode drops the audio track. So keep your original clip. Dragging the sound back in takes one step in CapCut, iMovie or DaVinci Resolve.

Run this before you upload

What the queue is actually doing while you wait

A photo swap runs the model once. A 10-second clip at 24 frames per second runs it 240 times. Every frame gets decoded, face-detected, landmark-aligned, swapped and blended. Then the whole thing is re-encoded into one MP4.

That is why the wait feels long compared to a still image. Frame count drives the clock, not clip length on its own. A 15-second clip at 60 fps has more frames than a 30-second clip at 24 fps, so it takes longer despite being half the duration.

Switching Frame Quality to High adds an enhancement pass on every frame, which stretches the job further. Most clips still land in about two minutes. A full 60-second clip can take several.

Photo, GIF or video: which tool your job needs

Three separate tools exist because the formats behave differently. Picking the wrong one wastes your daily allowance.

FormatFree-account ceilingOutput fileTypical waitBest for
Photo2 swaps a dayPNGSecondsSingle images, profile shots, memes
GIF1 a day, 5 MB, 150 framesLooping GIFUnder a minuteReaction loops for Discord, Slack, chat apps
Video1 clip a day, 20 s, 5 MBH.264 MP4About two minutesShort-form posts, edits, gifts, pre-vis
How many times the model runs per job, at each format's published free-account ceiling
Photo swap1 frame
GIF, free150 frames
Video, free (20 s)480 frames
Video, Pro (60 s)1,440 frames

Frame ceilings come from the published plan limits on each tool page. Video figures assume 24 frames per second, which is the standard rate for phone and web clips. Your own clip's frame count will differ if it was shot at 30 or 60 fps.

Swapping one still? The photo face swap tool finishes in seconds and skips the encoding step entirely. Making a looping reaction clip? The GIF face swap tool preserves the original frame delays and loop settings, which a video export cannot do.

Why face swaps fail, and what each failure looks like

This is the part most guides skip. The model needs a mostly frontal view to place face landmarks reliably. When it cannot find a face in a frame, that frame passes through completely unchanged.

Understanding that one rule explains almost every bad result. You do not get a smeared face. You get the original face back for a few frames, which reads on screen as flicker.

The seven common causes

What you seeLikely causeFix
Face flips back to the original during head turnsProfile angle. Landmarks cannot be found past roughly three quarters turn.Trim to the frontal section of the shot, or accept the flicker as a cut point.
Swap breaks whenever the subject moves fastMotion blur destroys the edges the detector needs.Choose a steadier clip. Handheld and running shots are the worst case.
Eyes look wrong or the swap drops in and outGlasses. Frames and reflections sit directly across the landmark points.Use a source photo without glasses, and a clip without them too.
Swap disappears when hair blows across the faceOcclusion. Hands, masks and hair all block detection the same way.Pick a clip section where the face stays clear.
Flat, pasted-on look at the jaw and templeLighting direction differs between the photo and the clip.Match them. A side-lit clip needs a side-lit source photo.
Soft or mushy swapped faceSource photo resolution is too low, or the face is too small in it.Reshoot or recrop so the face fills a third of the frame.
Everyone in the shot has the same faceSwap All Faces is on, and only one source identity exists per job.Turn it off for single-subject clips. Run the tool once per identity otherwise.

The multiple-face rule people miss

Each job accepts exactly one source face. There is no per-person targeting. Leave Swap All Faces on in a crowd scene and everyone gets the same identity. That looks strange fast. Want two different faces in one clip? Run two jobs, then composite them in an editor.

The edge fix that costs nothing

Seeing a hard line around the swapped face? Drop Blend Strength from 1.0 to about 0.85. Full-strength swaps are cleanest on still faces. Compression artifacts and motion blur in a moving clip make that boundary more visible, and a small blend softens the join without weakening the effect.

How to judge the output in 30 seconds

Watch the clip once at full speed. Then scrub through it slowly and look at three places only.

Check the jawline first, because that is where blending shows. Look for a colour band or a visible seam. Next, watch the hairline through any head movement. Last, count the flicker: pause on any frame where the original face returns, and note whether that moment matters in your edit.

A result that holds through the head turns is finished. One that flickers at a key moment usually needs a different section of clip, not different settings.

What the free account covers, and when Pro is worth it

Free accounts get one video swap a day, up to 20 seconds and 5 MB, watermarked. Finished clips stay available for 24 hours. That is enough to test the workflow and post something short.

Pro and Max lift that to 60-second clips at 100 MB, drop the watermark, allow up to 200 clips a day, and keep files while you subscribe. Each clip bills at 30,000 tokens, roughly a cent. Both plans start with a 7-day free trial, and current numbers sit on the pricing page.

Your first swap, in order

Pick a 10-second clip where the subject faces the camera and barely moves. Grab a front-facing portrait with even light. Run one swap at the default settings before you touch a single slider, so you learn what "normal" looks like on your footage.

Then change one thing at a time. Drop the blend to 0.85 if edges bother you. Raise Frame Quality only when the face is large in frame. No clip to work with yet? Make one with the AI video generator, then swap into it. More walkthroughs for the other tools live in the how-to guides.

More from the blog