WaveSpeedAI InfiniteTalk

Take a photo of a person, hand it a piece of audio, and get back a video of that person saying it, with the mouth and expressions synced to the speech. That's InfiniteTalk - a talking-head / lip-sync model that hit the scene in August 2025 and got the internet very excited, very fast. This node from razvanmatei-sf/razv-wavespeed runs it through the WaveSpeed AI API, so no weights, no GPU, just an API key and a bill.

The obvious use is dubbing and avatars - voiceover a character, localize a video into another language while keeping the original performance's energy. The tooltip even mentions prompt can steer scene, pose, and behavior while preserving audio sync, so it's not a frozen-face puppet: you can direct some of the motion. There's also mask_image, to say which person in a group photo should do the talking. That last one is the detail that makes this node feel properly built rather than hastily wrapped.

How it works

WaveSpeedAI Client → image (the face to animate) + audio (the speech track) → POST to the InfiniteTalk endpoint → poll → video_url out. The output is a hosted video URL, not a file - download it before it expires.

The inputs that actually matter:

• audio - an audio file URL. The tooltip says connect from an "Upload Audio node," and just like the image-upload situation, that node isn't registered in the current build - so paste a publicly reachable audio URL.
• image - the person to animate. Public URL.
• resolution - 480p or 720p, and the tooltip prices it out: $0.15 per 5 seconds at 480p, $0.30 at 720p. Yes, the cost is in the node. That's this pack being honest.
• prompt - optional instructions for scene/pose/behavior while keeping audio sync. This is the differentiator; use it.
• mask_image - optional, to select which person in a multi-person image animates.
• enable_base64_output - an optional toggle for getting the result as base64 instead of a URL. Leave it off unless you know why you need it.

seed (-1 = random) and enable_sync_mode (defaults off) round it out.

Install

Shared pack:

cd ComfyUI/custom_nodes
git clone https://github.com/razvanmatei-sf/razv-wavespeed
# restart ComfyUI

Or ComfyUI Manager → search "razv-wavespeed" → Install. Add a WaveSpeed AI key via the WaveSpeedAI Client node, config.ini (copy from config.ini.tmp, [API]), or a WAVESPEED_API_KEY env var.

Common issues

• It's priced per 5 seconds and video-shaped, so a minute of output is real money. The tooltip tells you the exact cost; believe it.
• URL inputs, and no upload node. The tooltips reference "Upload Audio"/"Upload Image" nodes that aren't in the current build. Host your files and paste the URLs.
• Face quality sets the ceiling. A small, side-on, or low-res face won't lip-sync convincingly. Use a clean, frontal, well-lit headshot.
• Sync fidelity varies with speech complexity - long silences, music, or multiple voices can confuse the alignment. Keep the audio track clean if you can.
• Ethics note, because this class of model needs one: don't put words in a real person's mouth without consent. It's that simple.