Companion episode — InfiniteTalk is the newest talking-avatar method we tested.
What InfiniteTalk is
InfiniteTalk (MeiGen / MultiTalk) turns a portrait plus an audio track into a lip-synced talking video, running locally in ComfyUI via the kijai WanVideoWrapper. Its headline feature is right there in the name: unlimited length — it uses a sliding-window approach to keep going as long as the audio does, and it generalizes surprisingly well (anime, even animals).
The stack
- Wan 2.1 I2V 14B base + the InfiniteTalk module
- a wav2vec audio encoder
- a step-distill LoRA so it samples in ~6 steps
- attention set to standard PyTorch (sdpa) unless you’ve built the faster attention kernels
It renders in windows — so a longer clip is genuinely several windows stitched by the pipeline, not one giant pass.
The honest verdict from our rig
It looks very good — the lip-sync is clean and the identity holds. But here’s the catch we hit on a 32GB card: the “infinite” part isn’t practical yet. Short clips (a few seconds) are stable; pushing toward a genuinely long render fell over — the full-length output tensor exhausted memory and crashed the app. So on consumer hardware today, it’s promising, not infinite.
If you’ve got more VRAM (or rent a big card), the ceiling moves. On 32GB, treat it as a high-quality short talking-clip tool, and watch this space — the approach is the right one.
Where it fits
A new and improving talking-head method worth having in your kit — especially if it generalizes to a style (anime) that other tools choke on. It’s not our daily driver yet, purely because we can’t push the length on our own hardware. Independent comparisons also rate it a notch below the more established full-body methods for natural motion over long runs — consistent with what we saw.
Honest verdict
Genuinely good lip-sync, genuinely exciting “infinite” idea — but until you can push the duration on your own card, it’s not quite infinite enough to be a workhorse. Promising. Test it; just don’t expect the full length on 32GB.