FaceFusion Is a Parts Assembler, Not a Method — A Map of Four Families of Face-Changing AI

I heard "I did it with FaceFusion" so often that I assumed FaceFusion was the name of a face-swap method. It isn't. It's a specific piece of software, facefusion/facefusion on GitHub. Not an academic term, not a method.

Face-changing AI splits into four families, and tools from different families aren't comparable. After reading this you'll be able to sort the tool, model and service names that show up mixed together in search results, and pick the family that fits what you want.

One line first: use only your own face or a face you have consent for, and label anything you publish as AI-altered (a deepfake). Everything below assumes that.

Terms first

Face swapping moves A's identity (who they are) onto B's face. That's FaceFusion's main job. Face reenactment moves A's expressions and motion onto B, the opposite direction. Deepfake is the popular umbrella term for all of it.

Swapping itself comes in two kinds. One-shot needs a single photo (inswapper, hyperswap, SimSwap). Subject-specific trains per pair of people (DeepFaceLab).

The four families

The interesting pair is A and C. Same input, one photo plus one video, opposite result.

When I tried A, the original actor's shaved head stayed in the output. By design, C can't have that problem. Instead the background and body come from the photo. I haven't run C myself yet, so that part is what the table predicts, not something I measured.

Which one, when

• Keep the original video's acting, lighting and camera, change only the face -> A

• You need new images of your face (profile picture, thumbnail) -> B

• Make a single photo talk or move -> C

• Change hair, clothes and body too -> D. You get a newly generated scene, not a remake of the original

"Which is better, FaceFusion or Runway?" isn't a real question. One is A, the other is D.

FaceFusion is a parts assembler

Inside FaceFusion you swap the model at each stage. Whatever combination you run, people call it "done with FaceFusion".

Beyond swapping, you can plug processors like lip_syncer, expression_restorer, age_modifier, frame_enhancer and background_remover in and out.

That skeleton (detect, align, swap, mask, composite, enhance) isn't unique to FaceFusion either. roop and Deep-Live-Cam have the same shape. A new paper usually contributes one swap model in the middle and borrows the stages around it. That's why SimSwap and HifiFace show up as options inside FaceFusion rather than as competitors. I took the stages apart one by one in why face swap can't change your hair.

The roots are InsightFace

Many family-A parts come from the InsightFace project: ArcFace, which turns a face into an identity vector (I used it as arcface_w600k_r50), the RetinaFace and SCRFD detectors, and inswapper_128, where one-shot swapping started.

Check the license. The InsightFace repo says the code is MIT, but the training data and models trained on it are for non-commercial research only, and for the inswapper series it asks you to contact them about licensing (checked 2026-10-19). A local experiment on your own machine and a public service are very different things.

If you want face swapping or avatars in a product and need to sort out which family and which licenses to go with, you can ask negabaro (contact).