What is InfiniteTalk?
InfiniteTalk is an audio-driven video model from MeiGen-AI. Its GitHub repository describes it as audio-driven video generation for "sparse-frame video dubbing", and its news list dates the release of the technical report, weights and code to August 19, 2025, the same day the paper was submitted to arXiv. The project page lists Meituan among the authors' affiliations, alongside universities and research institutes.
The README says the model syncs not only the lips but also the head, body and expressions to the audio, and that it builds on the Wan2.1 image-to-video model, with an audio encoder that reads the voice.
From a photo or from a video
InfiniteTalk works two ways. Its project page shows it dubbing an existing video, keeping the identity, background and camera movement while the person speaks the new audio, and its README says it can also animate a single image. The README lists 480p and 720p, and a separate checkpoint for several people. It also notes that from a single image, results stay good for up to about a minute, after which colors drift more.
On MakeMidas, InfiniteTalk animates one portrait (JPG, PNG or WebP) to a voice track of up to 15 seconds, at 480p or 720p. You can add a short direction for expression and movement. For longer videos, Kling AI Avatar 2.0 takes up to 5 minutes of voice; both are on the AI talking avatar page.
A script instead of a recording
InfiniteTalk speaks the audio you give it, so the voice is yours to choose. Record yourself, or turn a script into a voiceover with text to speech, download it, and upload it with the photo. For an AI spokesperson or AI actor who does not exist, make the portrait first in the AI image generator.














