Processors

The processors that do the work

Processors run per frame, in the order you list them with --processors. Each one targets a region (face, mouth, frame) with its own model selector and blend options.

face_swapper

face_swapper.py

Replaces the target face region with the source identity. The model choice is the single biggest lever on identity likeness vs. robustness. Models differ in latent resolution and how they encode the source identity (fixed reference, ArcFace embedding, or a dedicated face network).

Options

--face-swapper-pixel-boost - Upscale the swapped region in a second pass for visibly sharper output. (Added in newer releases; choices are resolution steps (e.g. 128/256/512/1024).)
--face-swapper-pixel-boost-blend - Blend the upscaled swap back over the original region. (Version-dependent range/default.)

Models

Model	VRAM	Why / when	Pros	Cons
inswapper_128	mid	The default workhorse. A 128px latent, InsightFace-style swap that runs from a source reference face and is widely regarded as the most identity-accurate general model. Default for most swaps; pair with an enhancer for crisp close-ups.	+ Best out-of-the-box identity likeness + Robust across angles/lighting + Most supported	- Soft at high zoom - 128px latent limits fine detail
simswap_256	low	Drives identity from the source face's ArcFace-style embedding rather than a reference frame. Good when you only have a single clean source. Single clean source image, modest hardware.	+ Good identity from one clean source + Lower VRAM than larger nets	- More sensitive to occlusion/pose - Smaller ecosystem tuning
simswap_512recent	mid	Higher-resolution sibling of simswap_256 - more latent capacity for detail at higher VRAM cost. Embedding-driven swap where you want more resolution.	+ Sharper detail than 256 + Good embedding-driven likeness	- Heavier on VRAM/throughput
ghost_256	low	A 'Ghost' face-swap network designed with occlusion tolerance in mind; often faster to run. Fast work, partially occluded or moving subjects.	+ Tolerates occlusion/cropped faces + Fast throughput	- Less identity fidelity than inswapper in clean conditions
uniface_256recent	mid	Facefusion's own UniFace network. A newer option that aims for a strong balance of identity and speed. Newer installs wanting a modern default swap.	+ Newer, actively maintained + Good identity/speed balance	- Not present in older releases

Run order is meaningful

Chaining multiple processors in one --processors list runs them sequentially on every frame. Facefusion processes each frame through the list, so put the swapper before the enhancer if you want the enhancer to restore the swapped face. Heavy multi-stage work is often better split into separate runs (see the encoding guide).