This ComfyUI workflow performs a video face swap by applying a single reference face image to every frame of a driving video, then recombining the frames into a final clip. The pipeline starts with VHS_LoadVideo to decode the source video into frames and LoadImage to bring in your reference face. A custom face-swap node (dc72113f-8276-4a5a-af12-85d6bec89ed5) drives the actual face transfer using the LTX 2.3 LoRA trained by @Alissonerdx. PrimitiveFloat and PrimitiveInt nodes expose key controls (such as effect strength, region size, or frame step), while ImageResizeKJv2 aligns the reference face scale to your target footage.

To stabilize results and keep the background intact, the workflow composites the generated face back onto each original frame with ReservedRegionFrameComposer. This node lets you define a consistent “reserved” region for the face and blend it with feathered edges, reducing temporal jitter and preserving hair, lighting, and scene motion. ComfyMathExpression is used to derive dynamic values (for example, region size from video resolution), keeping settings in sync with whatever you load. For quality checks, ImageConcanate can assemble side-by-side previews of original vs. swapped frames. Finally, the processed frames are encoded to a video using VHS_VideoCombine or CreateVideo, and the result is written out with SaveVideo.