AIFaceSwap: Swap a Face in One Photo, Step by Step
Two photos, no prompt, about 10 seconds. Here's how to run AI Face Swap V1 on each::labs, choose photos that swap cleanly and fix the results that don't.
By Eftal Yurtseven · Nov 16, 2026 · 9 min read

Most bad face swaps go wrong before the model runs. Someone picks a squinting selfie shot under a kitchen bulb, drops it onto a sunlit portrait taken from a three-quarter angle, and then blames the AI for the waxy result. AIFaceSwap can't invent the information you didn't give it. Choose the right pair of photos and the swap mostly takes care of itself.
This guide walks through AI Face Swap V1 from Aifaceswap on each::labs, one step at a time: which image goes where, what the model page actually lets you set, how to pick photos that swap cleanly, how to rescue a bad result, and where the model sits inside a bigger workflow.
What AIFaceSwap Does on each::labs
The model card describes it plainly: a face swap model that "automatically replaces the face in an image with another face while preserving expressions, lighting, and overall realism." It's an image-to-image model, it returns a still picture, and the page lists an average run time of around 10 seconds. That speed matters more than it sounds, because the real work here is trying a second or third face photo, not waiting.
Aifaceswap is a provider that does one thing, and the Faceswap family on each::labs currently holds a single model. No prompt box, no style presets, no sliders. That's a feature. When the only things you control are two photos, you know exactly where to look when something goes wrong.

Know Your Two Inputs Before You Upload
The form has two required fields, and the labels are where people trip. Here's what each one means, straight from the model's request schema:
Source Image (source_image) is the picture that gets edited. The schema calls it the "source/target image where the face will be replaced," and it's explicit that "the body, background, and all non-face elements remain unchanged." Think of it as the scene: the outfit, the pose, the setting you want to keep.
Face Image (face_image) is the face donor. The schema says the face "will be extracted and blended onto the source image, matching lighting, angle, and skin tone." This is the identity you're bringing in.
So the word "source" points at the scene, not the face. If your first result shows the wrong person on the wrong body, you swapped the slots. Both fields take JPG, JPEG or PNG files, and that's the whole configuration. There's no strength control and no seed to tune, which means every quality decision you make happens when you choose the photos.

Swap a Face Step by Step
Here's the full run in the each::labs playground, with nothing skipped.
Step 1: Open the model page
Go to AI Face Swap V1 and sign in. The playground shows the two upload fields on one side and the output panel on the other.
Step 2: Upload the scene as Source Image
Pick the photo whose body, clothing and background you want to keep. A single, clearly visible person works best. Crop away anything you don't need before uploading, especially other faces near the subject, so there's no doubt about which face gets replaced.
Step 3: Upload the identity as Face Image
Use a sharp, well-lit, front-facing photo of the person whose face you're bringing in. A passport-style shot is boring and ideal. The model page says it "performs best with high-resolution images featuring clearly visible, frontal or near-frontal faces," and that applies to both photos, but the donor face is where it shows first.
Step 4: Run and inspect at full size
Generate and open the output at 100 percent. Check the jawline, the hairline and the edges of the eyes, because that's where blending gives itself away. A thumbnail always looks fine. Zoomed in, the truth comes out.
Step 5: Change one photo at a time
If the result is close but off, keep the scene and try a different face photo, or the other way round. Changing both at once teaches you nothing. At around 10 seconds a run, three careful attempts take less time than one bad edit in a photo editor.
Pick Photos That Swap Cleanly
This is the section that decides your result. The model matches lighting, angle and skin tone, but matching works best when the gap is small to begin with.
Match the angle. A front-facing donor onto a front-facing scene is the easy case. If the scene shows a head turned in three-quarter view, find a donor photo turned the same way. The model page names profile angles among the cases where results get less accurate, so a full side profile is a gamble you'll usually lose.
Match the light. Soft daylight on both photos beats a perfect face shot under harsh overhead light. Watch the direction too: if the scene is lit from the left and the donor face is lit from the right, the shadows on the swapped face fight the rest of the picture. Low light is another case the page flags, so skip dim, grainy shots entirely.
Keep resolution honest. The face area in each photo should have real pixels in it. A face cropped out of a group shot and blown up is mostly guesswork, and the swap inherits that blur. Avoid heavily compressed images forwarded through three chat apps.
Clear the occlusions. Sunglasses, a hand on the chin, a microphone, hair across one eye: anything covering the face confuses the extraction. Obscured faces are the first item on the model page's list of weaker cases. A fringe is usually fine. Mirrored aviators are not.
Match expression roughly. The model preserves expressions, so a wide laugh in the donor photo on a calm scene can look strange. You don't need an exact match, but a neutral or slight smile on both sides is the safe default.

Fix the Common Bad Results
When a swap misses, the failure usually tells you which photo to replace.
The face looks pasted on. Visible seams or a face that seems brighter than the neck point to a lighting mismatch. Find a donor photo shot in similar light, or pick a scene with softer, more even lighting. Cropping the donor tighter around the face can help too, since there's less stray background competing with the face itself.
The features are smeared or soft. That's resolution or compression. Replace the blurry photo with the original file rather than a screenshot. If the scene photo itself is old or degraded, restore it first: our GFPGAN walkthrough covers how to sharpen faces in old photos before you use them as a base.
The face is warped or stretched. The angles don't agree. Match the head turn more closely, and avoid donors shot from far above or below. Selfies taken at arm's length with a wide lens distort the nose and chin, and those distortions travel with the face.
The wrong face got replaced. The scene had more than one person. Crop it down to the one subject before uploading, then composite the result back into the full image if you need the group.
The person on the result isn't who you expected. Check the slots. The identity belongs in Face Image, the scene in Source Image.
If you've done all of that and still want a different look, each::labs also runs Eachlabs Face Swap, which takes a Source Image and a Target Image, and Face Swap, which its page says is optimized for single-face swaps. Running the same photo pair through a second model is a fair test, and it often tells you whether the problem is the model or the photos. Usually it's the photos.
Use Face Swaps With Consent
A face is someone's identity, and putting it on a body that isn't theirs is something you should only do with their permission. The model page itself calls the model "suboptimal for complex, unauthorized, or deceptive uses," and that's the right line to hold. Swap your own face, swap the faces of people who've said yes, and use models or talent who've signed off on the use when the work is commercial. Don't make images meant to pass as real evidence of something that didn't happen, don't put a real person into intimate or humiliating scenes, and don't use the face of a public figure to imply an endorsement they never gave. If you publish a swapped image, label it as edited when there's any chance someone could take it as a real photo. None of this makes the tool less fun. It just keeps the fun from landing on someone who didn't ask for it.

Where AIFaceSwap Fits in a Workflow
A face swap is rarely the whole job. It's usually one step between a shoot you already have and the asset you actually ship, and that's where each::labs AI Flows come in: you chain models so the output of one feeds the next without moving files by hand.
A typical still-image chain looks like this. Restore or upscale an old or soft scene photo, swap the face with AI Face Swap V1, then pass the result to an edit model for final touches like background or color. Product teams use the same idea for e-commerce visuals and campaign templates, where one approved model shot gets reused with different consenting faces. The model page lists exactly those cases: ad visuals for marketers, portrait concepts for designers, "swap my face onto this group photo template" features for app developers.
If you want an edit model to handle the swap itself, there's a ready-made flow, Nano Banana Pro Face Swap. It runs Google's Nano Banana Pro Edit in one step, and its description says it keeps the source face's identity, proportions and expression while matching the skin tone, lighting and shadows of the target image. It's a good second opinion when a scene has tricky lighting.
For video, the still-image model isn't the tool. PixVerse Swap replaces a subject in an existing clip using a reference image, and our PixVerse Swap guide walks through it. And if you're building face swap into your own product rather than using the playground, the face swap API guide for consumer apps covers uploads, moderation and consent checks at the code level.
A Quick Checklist Before You Hit Run
Before every swap, look at the two photos side by side and ask five things. Is the scene in Source Image and the identity in Face Image? Do both heads point roughly the same way? Is the light coming from a similar direction with similar softness? Is each face sharp at full size, with nothing covering it? And does the person in the donor photo know what you're doing with their face? Five yeses and AIFaceSwap tends to give you something you'd keep on the first or second try.
Frequently Asked Questions
Is AIFaceSwap a prompt-based model?
No text goes in at all. AI Face Swap V1 takes exactly two inputs, Source Image and Face Image, so every quality decision happens when you pick the photos. That's why lighting, angle and resolution matter so much more here than wording ever could.
Which photo is the source and which is the face?
Think scene versus identity. Source Image is the picture that gets edited, and its body, background and other elements stay as they are. Face Image is the donor: the face extracted from it is blended onto the scene.
What file formats does AI Face Swap V1 accept?
JPG, JPEG and PNG for both inputs, according to the model's request schema. Upload the original file rather than a screenshot or a compressed copy, because the swap can only be as sharp as the face you give it.
Can I swap faces in a video with this model?
This one is still images only. For clips, PixVerse Swap on each::labs replaces a subject in an existing video using a reference image, and it pairs well with a still swap when you need matching photos and footage of the same character.



