FAL Mirage Avatar X Reference To Video (Soze)

Most video nodes ask you to describe motion. This one asks you to plug in a microphone. Mirage Avatar X is an audio-driven avatar model: you feed it a clip of someone talking (3–180 seconds), and it produces a video of an avatar lip-syncing to that audio. It's the "face that moves to a voice" job that otherwise means wiring up a talking-head model, a lip-sync model, and a prayer - here it's one node and one FAL call.

The inputs

The required input is audio, and it's an AUDIO object, not a file path - wire it from whatever audio source your workflow has (a TTS node, a loaded clip, an ElevenLabs output). That audio drives everything: the speech, the mouth movements, the timing.

Then you pick who's talking. Three ways, in order of precedence:

• video_reference - a reference video of a person. Highest control; the avatar looks like them and moves like them.
• image_reference - a single reference photo (9:16 or 16:9). Cheaper than a video ref, and usually enough for a static portrait avatar.
• avatar - a stock avatar dropdown (None, Ayesha, Farhan, Giulia, Jasmine, Luke, Maya, plus 16:9 variants). Instant, no reference media needed.

If you connect both a video and an image reference, the video wins. If you connect any reference at all, it overrides the stock avatar selection.

How it works

Same FAL wrapper machinery as the rest of the Soze pack: your audio (and optional reference media) uploads to FAL's CDN, the node POSTs to mirage-api/avatar-x/reference-to-video, waits synchronously, then downloads the result to your output folder as a VIDEO object. Outputs are video, video_url, and status.

The realistic expectations: this is a hosted avatar product, so the output looks like an AI avatar - clean, presentable, and a little uncanny at close range. It's built for narration, product explainers, and talking-head content, not for winning an acting award. For that use it's impressively turnkey.

Install and key

From the Soze pack ("Quality of Life Nodes for ComfyUI"). ComfyUI Manager → search "Soze", or:

cd ComfyUI/custom_nodes
git clone https://github.com/SozeInc/ComfyUI_Soze.git
pip install -r ComfyUI_Soze/requirements.txt

Restart, then set FAL_KEY as an environment variable or in ComfyUI_Soze/config.ini:

[API]
FAL_KEY=your_key_here

Key at fal.ai/dashboard/keys; don't commit the config file.

Common issues

• "No audio" error. Make sure an AUDIO object is actually connected - this node has no text-to-speech mode; it only animates audio you supply.
• Avatar looks nothing like your reference. Check which input is winning - a connected video_reference silently overrides image_reference, and either overrides the stock avatar.
• Lip-sync drifts on long clips. This is an avatar service, not a forensic lip-sync tool. Keep clips under a couple of minutes.