InfiniteTalk
Audio-driven model that makes unlimited-length talking video from an image or re-dubs an existing video.
Turn an image plus audio into long talking video, or re-dub a clip with matching body motion.
InfiniteTalk is a sparse-frame video dubbing model from MeiGen-AI that generates talking video of unlimited length from an image or an input video. It suits developers and researchers with a strong NVIDIA GPU.
What it is
Given audio, it synchronises lips and also head posture, body movement and expressions, in image-to-video or video-to-video mode. The README builds on the Wan2.1 480P base model and offers a Gradio demo, ComfyUI branch and low-VRAM mode.
Who it's for
Suits technical users who want long avatar clips without per-minute fees. It does not suit anyone expecting a studio with voices, templates or a script editor, because it only consumes audio you provide.
Key features
Image-to-video
Animate one reference image from an audio clip.
Video-to-video dubbing
Re-dub an input video, syncing head, body and expressions.
Unlimited length
The README describes unlimited-length generation.
Low-VRAM mode
Set num_persistent_param_in_dit to 0 for very low VRAM.
Multi-GPU inference
Run with torchrun across several GPUs.
Gradio and ComfyUI
A Gradio demo and a ComfyUI branch are provided.
Up and running in 3 steps
Our honest verdict
What's great
- Apache-2.0 code and weights, with the README granting freedom to use generated content.
- Dubs both lips and body motion, not only the mouth.
- Documents low-VRAM and multi-GPU modes for different hardware.
- Active repo with ComfyUI and Wan2GP community integrations.
Where it falls short
- You need an NVIDIA GPU, PyTorch 2.4.1 with CUDA 12.1, xformers and flash-attn compiled correctly.
- Base Wan2.1-I2V-14B weights and extra audio models must be downloaded separately; they are large.
- No voices, editor or text-to-speech: you supply the audio.
- Low-VRAM mode trades away speed.
Pricing & license
Free under Apache-2.0. The README lists no paid or hosted plan, so the only cost is the GPU hardware or rented GPU time you run it on.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I use InfiniteTalk commercially?
The README says the models are under Apache-2.0 and that you keep rights to your generated content, subject to the licence terms.
How much VRAM does it need?
The README does not give one number; it documents a very-low-VRAM mode and multi-GPU runs, and community Wan2GP support targets low-VRAM cards.
Does it make the voice too?
No. It takes audio you supply and animates the video to match it.
