Expression Video

Here's the neat one. You've got a single still portrait and two expressions you want it to travel between - say, neutral to a big grin, or eyes-open to a wink. ExpressionVideoNode generates the in-between frames and hands you a smooth little animation of the face morphing from one to the other. No driving video, no acting, no camera. Just a start pose, an end pose, and a number of frames to fill the gap. The README flags this as one of the pack's newer additions - "you can now create interpolation animation between two expressions."

It pairs directly with the Expression Editor. That's where the two expressions come from: dial a face in the editor, grab its expression_json output, dial a second face, grab that one too, and this node tweens between them.

How it works

Every expression the editor produces is really a point in LivePortrait's motion space - a bundle of numbers for pose, eyes, and mouth. Given a start point and an end point, ExpressionVideoNode interpolates a path between them and re-renders the source face at each step along the way. Because it's warping the same still through the same engine, every frame is the same person in the same lighting; only the expression slides. The result is a clean tween rather than anything that had to be filmed.

The inputs and outputs that matter

• src_image (IMAGE) - the one still face that gets animated.
• from_expression (STRING) and to_expression (STRING) - the start and end poses, as the expression_json strings produced by ExpressionEditor_. These are the whole point; without a pair of them the node has nothing to tween.
• interpolations_num (INT, default 1) - how many frames to generate between the two poses. This is your knob for animation length and smoothness. One frame gives you a jump; crank it up for a longer, silkier morph. (The max is absurdly large, so use common sense - dozens for a smooth beat, not thousands.)
• interpolation_type (enum: linear / nearest / cubic, default cubic) - the easing curve. cubic eases in and out for a natural feel and is the right default; linear is a constant-speed morph; nearest snaps and is mostly there for completeness.

The output is frames (IMAGE) - the generated sequence as an image batch. Wire it into a video-save or combine node (Mixlab's, or any frames-to-video node) to write it out as an actual clip.

How to install it

ComfyUI Manager: search comfyui-liveportrait, install, restart. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/shadowcz007/comfyui-liveportrait

then restart ComfyUI. The LivePortrait and InsightFace weights come from the Google Drive / Baidu links in the README and go under ComfyUI/models/liveportrait and ComfyUI/models/insightface. Already running comfyui-mixlab-nodes? Then the dependencies are handled.

Common issues

The most common trip-up is the two expression inputs: they have to be real expression_json strings from the Expression Editor, not hand-typed or left empty. Build the editor → this-node chain first and confirm the JSON is flowing. After that it's the pack's usual suspects - missing model files in models/liveportrait (they don't self-download), and InsightFace being InsightFace on install (a missing C++ build toolchain historically, much smoother since InsightFace 1.0). Keep interpolations_num reasonable; a giant value doesn't make it smoother past a point, it just makes it slow. And the perennial license note: the InsightFace weights underneath are non-commercial, which carries over to whatever you export.