Open-Source Real-Time Avatar Models by the Numbers: Speed, Hardware and Licenses

What counts as live

A live talking-head avatar has to make video frames at least as fast as they are played. LiveAvatar's authors put the bar at 24 frames per second and say plainly that their "real-time" means "generation throughput exceeding playback speed, not end-to-end conversational latency".¹³ LiveTalking, an open framework that wires these models into a video call, uses 25: its inference and streaming frame rates "must be >= 25 for real-time performance".³¹

Speed alone isn't enough for a conversation. The other figure is how long the first frame takes, since the viewer waits that long before the face moves. The newer streaming models report it: 0.33 seconds for LiveTalk, 0.87 for SoulX-FlashTalk, 0.94 for Hallo-Live and 1.21 for LiveAvatar.^15192813 Ant Group's Ditto gives a real-time factor (time to generate divided by the length of the clip) of 0.895 for streaming a talking head, with 385 ms to the first frame, on one A100.⁵ And the avatar is only the last stage of a live pipeline: speech recognition, a language model and text-to-speech come first, and each adds its own latency.

Three ways to animate a face

The oldest approach in the table edits only the mouth. Tencent Music's MuseTalk repaints a 256×256 face region of an existing video to match new audio, in a single step in latent space; its makers stress that it is "NOT a diffusion model" and report "30fps+ on an NVIDIA Tesla V100".¹² Portrait-animation models move the whole head instead. Kuaishou's LivePortrait warps one still image to follow a driving video, at 12.8 ms per frame on an RTX 4090.³ Ditto adds a diffusion model that turns audio into head and lip motion, then renders it.⁵ Alibaba's LiteAvatar drives a set of prebuilt 2D avatars from audio and, its makers say, "can run in 30fps on only CPU devices".⁷

A second family works in 3D. Alibaba's LAM builds an animatable 3D Gaussian head from one photo in a single pass; its paper reports 280.96 frames per second rendering on an A100 and 26 on a Xiaomi 14 phone through WebGL.⁸ (Its README quotes higher figures, 562.9 fps on an A100 and "110+FPS" on the same phone; we use the paper's.⁹) NVIDIA open-sourced Audio2Face-3D on September 24, 2025: models, an SDK and a training framework.¹¹ It turns audio into motion for a 3D character's skin, jaw, tongue and eyes, and the v3.0 model, with 1.80×10⁸ parameters, ran at 3,269 frames per second on one RTX 4090 in NVIDIA's own test.¹⁰

The third family is new. From December 2025, research teams began distilling large video-generation models into "streaming" ones that generate a few frames at a time, conditioned on what came before, in a handful of denoising steps: four for LiveAvatar, SoulX-FlashTalk and LiveTalk, one for AvatarForcing.^13191524 LiveAvatar, for example, "distills a pretrained bidirectional model into a causal, few-step streaming one".¹³ They generate whole video frames rather than a face region, and several are built to stream for hours: LiveAvatar reports stable generation "exceeding 10000 seconds".¹³ Eight of the 15 models in the table are of this kind, all released between December 2025 and April 2026. (Person-specific methods such as SyncTalk, which trains a model of one person from a four-to-five-minute video, also reach real-time rendering but aren't one-shot, so they're left out.³⁷)

Six of the eight streaming models start from Wan

Alibaba's open-weight Wan video models are the base for most of the streaming avatars. LiveAvatar is built on Wan2.2-S2V-14B; SoulX-FlashTalk's architecture "derives from WAN2.1-I2V-14B"; SoulX-LiveAct is "14B Wan2.1 + 4B audio module"; SoulX-FlashHead is built "upon the Wan2.1 T2V (1.3B) architecture"; AvatarForcing lists Wan2.1-T2V-1.3B as its base model; and LiveTalk is distilled from "a multimodal variant of Wan2.1".^141927212515 Wan is released under Apache-2.0, as are these derivatives; LiveAvatar's README notes that "the Wan model (Our base model) is also released under the Apache 2.0 license".¹⁴

Two makers account for six of the 15 models. Alibaba teams released LiteAvatar, LAM and LiveAvatar, and Soul AI Lab released SoulX-FlashTalk, SoulX-FlashHead and SoulX-LiveAct between January and March 2026.^7914202227 Kuaishou, which makes the Kling video models, built LivePortrait, and members of its Kling Team co-wrote AvatarForcing.³²⁴

The hardware gap

Model size sets the bill. The two 14B models hit their stated speeds only on a server node: LiveAvatar reaches 45 FPS on five H800s with four sampling steps, and SoulX-FlashTalk 32 FPS on eight.¹³¹⁹ On one GPU they fall well short of playback. LiveAvatar's authors say their distilled model "reaches only 5 FPS on a single GPU", and SoulX-FlashTalk's README says "Real-time inference speed can only be supported on 8xH800 or higher graphics cards", with a single card needing more than 64 GB of memory.¹³²⁰ SoulX-LiveAct gets its 18B model to 20 FPS on two H100 or H200 GPUs using FP8 precision, and reports 6 FPS on a single RTX 5090.²⁷

The 1.3B models fit on one card. SoulX-FlashHead's Lite variant claims 96 FPS on a single RTX 4090, enough for "3-concurrent real-time(25+ FPS) streaming"; its larger Pro variant manages 10.8 FPS on one RTX 4090 and needs two RTX 5090s for real time.²² LiveTalk reports 24.82 FPS and AvatarForcing 34 ms per frame, each "on a single GPU" whose model the papers we read don't name.¹⁵²⁴ LiveAvatar's own figures have moved: its first paper said 20 FPS on five H800s before version 1.1 raised it to 45, and the latest version of the paper names the GPUs as H800 in its arXiv abstract but H100 in its body.¹³

What the licenses allow

Most of the table is under permissive licenses: Apache-2.0 for nine models, MIT for LivePortrait, MuseTalk's code and Hallo-Live, and NVIDIA's Open Model License, under which the Audio2Face-3D models "are commercially usable".¹² MuseTalk's README says its trained model is "available for any purpose, even commercially".¹ Four carry caveats in their own license text:

• LivePortrait is MIT, but the face-detection models it bundles from InsightFace "are for non-commercial research purposes only", so its license tells commercial users to "remove and replace InsightFace's detection models".⁴
• SoulX-FlashHead's Pro variant is Apache-2.0 throughout, but the Lite variant uses the LTX-Video 0.9 VAE, which keeps its Open-RAIL-M license.²³
• AVTR-1's weights allow commercial use only by organizations with annual revenue below US$10 million, and its renderer and streamer permit "no commercial use of any kind" without a separate license, so it is source-available rather than open source.³⁰
• LiteAvatar's code is MIT, but we found no license stated for its avatar gallery's weights.⁷

Outside the table, older and larger avatar models are often non-commercial. Wav2Lip "can only be used for personal/research/non-commercial purposes"; Tencent's Sonic says "our license is non-commercial"; and Tencent's HunyuanVideo-Avatar license "does not apply in the European Union, United Kingdom and South Korea".³⁴³⁵³⁶

From model to live conversation

A model makes frames; a product also needs speech-to-text, a language model, text-to-speech and a video stream. Three open projects assemble those parts:

• LiveTalking (Apache-2.0) streams MuseTalk, Wav2Lip or ER-NeRF avatars over WebRTC or RTMP. Its README lists MuseTalk at 42 FPS on an RTX 3080 Ti and 72 on an RTX 4090, and a 256-pixel Wav2Lip at 60 FPS on an RTX 3060.³¹
• OpenAvatarChat (Apache-2.0), from Alibaba's HumanAIGC engineering team, plugs LiteAvatar, LAM, MuseTalk or SoulX-FlashHead into a modular voice chat and states a "~2.2s average response time", without giving the hardware.³²
• Duix-Mobile is an on-device SDK for Android and iOS that states a response latency "under 120ms" on a Snapdragon 8 Gen 2. It ships four public avatars; custom avatars aren't part of the open release, and its community license requires a commercial license above 1,000 monthly active users.³³

Left out of the table

Many well-known open avatar models aren't real-time by their own documentation. EchoMimic V2 takes "~50s/120frames on A100 GPU" even after acceleration; MeiGen's InfiniteTalk and Tencent's HunyuanVideo-Avatar make no real-time claim; and Wav2Lip, the 2020 lip-sync model that LiveTalking still offers, doesn't claim one either.^38393634 Others claim real time but haven't released weights: Kuaishou's MIDAS and the TalkingMachines paper, for example, and Omni-LiveAvatar's repository says "Code and checkpoints are not released yet".⁴⁰⁴¹⁴²

Consent and safeguards

Most models here animate a face from a single photo or a short video, so the projects' own safety statements matter. AVTR-1's license forbids using it "to impersonate or attempt to impersonate (e.g. deepfakes) others without their consent".³⁰ LiveAvatar's paper encourages "access control and watermarking to prevent malicious use", SoulX-FlashHead's supports "invisible watermarking mechanisms", and LivePortrait's README discusses the deepfake risk directly.¹³²¹⁴ MuseTalk's and PersonaLive's READMEs disclaim responsibility for misuse.¹¹⁸ MetaPrompt covers avatar tech only for consenting people and doesn't publish step-by-step guides to animating a real person's face.

Reading the numbers

No independent benchmark yet measures these models side by side on the same hardware, so every speed here is Vendor-stated. The figures differ in GPU, resolution (from a 256×256 mouth region to 832×480 full frames), step count and what is timed: Audio2Face-3D's 3,269 fps is facial motion for a 3D engine to render, not video. A higher number in the table doesn't mean a faster live avatar on your hardware.