Video Face Swap · Video Edit

Faces in motion are a different problem

An image swap only has to be right once. A video swap has to be right in every frame, and - much harder - consistently: the same face, jittering as little as possible, through every turn of the head and change of expression, in a clip that already has its own lighting, compression and motion blur. That's where local swappers historically fall apart, and it's why video identity work has usually meant per-frame post-process hacks with temporal smoothing bolted on top.

This node is the one-call version: a video and a reference photo in, the facial features carried through every frame, with hair, movement, framing and background left as shot. $0.026 per second, so a 10-second clip is about 26 cents.

How it works

The pack uploads the video (90 MiB ceiling) and your reference image, calls spicyapi/face-swap-video-v1/video-edit, and returns the processed MP4 as a ComfyUI VIDEO. There's no analysis pass to run first - unlike the character- and head-swap models, this one does its own detection inside the single call, which is a real convenience and the reason it's the cheapest per-second video swap of the four.

Mechanically it's the ArcFace-embedding family again, the same recognition backbone under IP-Adapter FaceID, InstantID, PuLID and ReActor - hosted instead of installed. If you've fought InsightFace's build requirements locally, the appeal is obvious.

Inputs

video_url (VIDEO) - the target clip. The face of the person in it gets replaced. Load it with the core Load Video node.

reference_image_url (IMAGE) - the reference face, carried onto the person in the video. One clean frontal shot beats a folder of angles.

style - realistic (default), beautify, lossless. Same three finishes as the image swaps: realistic keeps natural skin tone and texture, beautify smooths and brightens, lossless holds the most facial detail. For video, realistic is the safe default - beautify on compressed footage tends to look waxy, and lossless amplifies whatever grain the source already had.

face_mapping (JSON) - multi-person mapping, a list of objects with "reference_face_index" and "target_face_index", 0-based in detection order. Leave it empty for the single-face case.

face_enhance (default off) - an enhancement pass on the swapped face. On video this is the higher-risk switch: a per-frame generative touch-up can flicker. Try it on a short trim before you commit to a long clip.

Outputs: video (VIDEO, to Save Video), task id (STRING). There's no seed field on this node, so runs aren't reproducible - budget accordingly.

Install

Manager → search SpicyAPI, or:

cd ComfyUI/custom_nodes
git clone https://github.com/spicyapi-ai/comfyui-spicyapi

Restart. No dependencies, no local model - the pack's dependency list is empty and all inference is server-side, so this works on a CPU-only box. Python 3.10+, a ComfyUI with the V3 node API, a key in Settings > SpicyAPI (or SPICY_API_KEY), then one restart so the node list builds from your account.

Traps

• Per-second pricing on a clip you'll iterate. It is very easy to burn a few dollars on one shot: test on a 3-second trim, and set Max cost per run to a number that doesn't hurt.
• Reference or target with more than one face. Without face_mapping, detection order decides. Run once, look, then map.
• Hard profiles and heavy occlusion. Turning heads, hands over faces, dark glasses - the swap degrades; that's the model's judgement, not a setting you can raise.
• Resolution and codec come back as served. Length, motion and the audio track are kept, but don't expect a bit-identical re-encode.
• No SpicyAPI key is set - key goes in Settings, never in the node, and never into a workflow file or the PNG metadata of your output.
• Cancelling doesn't stop the job. A cancelled run that completes server-side is charged, and the result waits in your console for up to 14 days.