Just tried out InfiniteTalk – it makes static portraits talk, and it's way more stable than MultiTalk.

Spent two whole nights on this, but I finally got the InfiniteTalk Video2Video workflow running in ComfyUI. Time to share some real-world tips.

This thing is an audio-driven lip-sync model, and the biggest selling point is no video length limit—feed it an audio clip, and the lip movements track super accurately, plus the head and body motions are pretty natural. I used MultiTalk before, and the hands would constantly twitch or the face would mess up; InfiniteTalk is way more restrained in that department, way fewer distortions.

You need six core models to download. The two fp16 diffusion ones are the biggest VRAM hogs—Wan2_1-InfiniTetalk-Single and wan2.1_i2V_480p_14B both have to go in their respective folders. There’s also a Chinese wav2vec2 that auto-downloads, so no manual fuss.

VRAM-wise, I gotta warn you: the official minimum says 24G. My 4090 barely handles 480p, and the moment I try 720p it’s straight OOM. If you want higher res, you’re looking at 48G+. If you’re hitting VRAM limits, turn on low mem load and switch to the fp8 version—that’ll save your ass.

Prompts can be dead simple, but the input audio quality needs to be top-notch. First time I used a compressed mp3, and the lip sync came out blurry as hell.

480p? Are you serious? That’s gonna lose all the facial details for sure.

24G is really just the bare minimum tbh, when you actually start running stuff you still gotta stack that VRAM.

The lip-sync on this is way better than old-school stuff like wav2lip.

Hey, quick question — for vertical videos, do I need to manually adjust the input dimension?

The fact that hands don’t break is already enough reason for me to switch. MultiTalk’s hands are a total nightmare.

The audio quality requirement part is way too real, I’ve been burned by that too.

OP, can you drop a finished piece so we can see? Just looking at the params doesn’t really give me a sense of it.