What Are ID Transfer and FaceSwap?
ID transfer is the people-specific branch of subject transfer. Most methods focus especially on keeping the face consistent (personally, I wish they'd treat hair and body as part of the ID too).
FaceSwap has been around since before generative AI. The classic approach puts a "mask" of the reference person's face over the target person's face, but if you use ID transfer to inpaint just the target's face, you get a much more flexible FaceSwap.
LoRA
LoRA is a way to "teach a model to draw things it couldn't draw before," and you can train it on images of a person too.
If you can live with the downside of having to train it, its flexibility and stability are top-notch.
IP-Adapter Family
There are several offshoots that take the IP-Adapter technique covered in subject transfer and specialize it for ID.
IP-Adapter (Original)
IP-Adapter is an adapter that adds "image-based conditioning" to existing text2image models.
It extracts feature vectors from an image and injects them into the UNet so they show up in the output. It can transfer people, but it's mainly for roughly carrying over a subject or style.
IP-Adapter-plus-face
The official face-specialized IP-Adapter model.
IP-Adapter-FaceID
The IP-Adapter-FaceID models work with a face recognition model (InsightFace) to lock in identity more strongly.
FaceID uses "Face ID embeddings from a face recognition model" instead of CLIP image embeddings, and adds a LoRA to further improve ID consistency.
The improved FaceID-Plus uses both "Face ID embeddings for identity" and "CLIP image embeddings for facial structure," getting a good likeness while keeping the face shape and structure stable.
InstantID
Strictly speaking it isn't part of the IP-Adapter family, but like those, it's an ID transfer method you can use just by plugging in an extra adapter.
It combines an IP-Adapter-style image adapter with an extra network (IdentityNet) that uses ID embeddings and facial landmarks from a face recognition model. This gives a better balance between keeping the identity from a single face photo and following text edits.
Personally, I think InstantID is the best option for SDXL.
PuLID-FLUX / InfiniteYou
Over time, the main arena for base models moved from SDXL to Diffusion Transformers (the FLUX family). ID transfer methods built for Flux.1 followed.
PuLID-FLUX is an ID-focused customization method built on FLUX1-dev. Thanks to tricks like Lightning T2I, you can change the art style while keeping the identity from just "reference face + text," with no extra training.
InfiniteYou is an identity-preserving framework built on FLUX-family Diffusion Transformers. With a module that injects identity features into the DiT itself (InfuseNet) and a multi-stage training strategy, it aims to improve likeness, text alignment, and image quality all at once.
Instruction-Based Image Editing and ID Transfer
Since they can do subject transfer, it goes without saying that "instruction-based image editing models" work for ID transfer too.
FaceSwap
FaceSwap replaces the target person's face with the reference person's face.
Classical FaceSwap (ReActor / InsightFace based)
Rather than ID transfer, the idea is closer to "making a mask and putting it on."
It uses face detection and landmark detection to estimate the face's position, angle, and outline, aligns the source and target faces with affine transforms and the like, then replaces the target's face with the source face using masks and blending.
The best-known option in ComfyUI is ReActor (InsightFace-based FaceSwap), but it's also a shaky technology to rely on; its repository has been taken down over ethical concerns, for example.
Face Inpainting using ID Transfer
Here you mask the target person's face and inpaint it, using ID transfer to redraw it as the reference face.
Back when ID transfer was still immature, one approach was to FaceSwap with ReActor first, then refine with ID transfer inpainting (it's one of the most popular workflows I've made).
Refining with InstantID after Face Swapping with ReActor
