Talking-head and lip-sync video from audio with open models: HuMo vs MultiTalk vs SoulX-FlashTalk vs SoulX-FlashHead (inputs, length, GPU needs, licences)

Open models that make a person talk or sing from audio: HuMo, MultiTalk, SoulX-FlashTalk and SoulX-FlashHead This compares what four open audio-driven video models' cards state. Lip-sync quality and speed were not tested here; quality claims are the authors'. What each is for - HuMo (17B and 1.7B): human video from text plus audio (mode TA), or text plus a reference image plus audio (mode TIA). Text plus image without audio is advertised but its inference is not released. - MultiTalk: audio-driven conversational video for one or several people; the authors say it also handles cartoon characters and singing and claim state-of-the-art lip synchronization. It ties each voice to the right person by locating that person's region in the reference image. - SoulX-FlashTalk-14B: real-time streaming of audio-driven avatars with no fixed length; the authors credit InfiniteTalk and Wan as the base they built on. - SoulX-FlashHead-1.3B: real-time streaming talking heads in two released variants, Lite and Pro. Length and resolution stated - HuMo: 480P and 720P for 17B; trained on 97-frame clips at 25 FPS, and the authors warn longer generations may degrade. - MultiTalk: 480p and 720p at arbitrary aspect ratios, up to 15 seconds. - FlashTalk and FlashHead: described as infinite streaming; their cards give no resolution figures. Hardware stated - HuMo-1.7B: listed as lightweight on a 32G GPU, generating a 480P video in 8 minutes there, per the authors; no figure for 17B. - MultiTalk: none on this card; installation and usage are in its GitHub repository, not inspected here. - SoulX-FlashTalk-14B: one GPU needs more than 64 GB of VRAM, and real-time speed needs 8x H800 or better. A 4-GPU version and a model for consumer GPUs such as the 4090 are announced as coming soon. - SoulX-FlashHead-1.3B: Lite reaches 96 FPS, or three concurrent real-time (25+ FPS) streams, on one RTX 4090. Pro runs at 10.8 FPS on one RTX 4090 and in real time on two RTX 5090s with SageAttention. Audio encoders and setup - HuMo: openai/whisper-large-v3, plus optional huangjackson/Kim_Vocal_2 to remove background noise. - SoulX-FlashTalk: TencentGameMate/chinese-wav2vec2-base. SoulX-FlashHead: facebook/wav2vec2-base-960h. - Both SoulX models: Python 3.10, torch 2.7.1 from the CUDA 12.8 wheels and flash_attn 2.8.0.post2. - MultiTalk: the card hosts weights only. Licences - HuMo, MultiTalk and SoulX-FlashTalk: card metadata declares Apache-2.0, and MultiTalk's card text says the same. - SoulX-FlashHead: metadata says "other". Its component list gives Apache-2.0 for the DiT weights, the Wan 2.1 VAE (Pro) and wav2vec2, but Open-RAIL-M for the LTX-Video 0.9 VAE used by Lite. Pro alone stays under Apache-2.0, and redistributing Lite must keep the Open-RAIL-M terms. Which to try (curator suggestion, untested) - One consumer GPU and real-time output: SoulX-FlashHead Lite, keeping its Open-RAIL-M VAE terms in mind. - Several people talking in one scene: MultiTalk. - Text-directed human video with audio and an optional reference image: HuMo in TIA mode. - Data-centre GPUs and long live streams: SoulX-FlashTalk-14B. Rights Animating a real person's face or voice needs that person's permission. None of these cards grants likeness rights, and this record grants no clearance. Detailed records: HuMo art_01M61VA3CSR5BWA1ETZJ9D87H3; MultiTalk art_01M5ZHE13CX5ZWZ3V5SCP4BGFF; SoulX-FlashTalk-14B art_01M5Z90HQ8Y8Y76B0EK2BVJ55D; SoulX-FlashHead-1.3B art_01M5ZECHQH5YQH8WB577MZ9YE6. Card statements only; lip-sync quality, speed and memory were not tested here. MultiTalk usage and all GitHub repositories were not inspected. Component licences other than those stated on the cards were not reviewed; no likeness clearance. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/bytedance-research/HuMo/raw/3a4a1610d399a5cbb932d54dc229944029803ff7/README.md",
      "note": "Retained source snapshot observed 2026-10-28; TEXT.txt lines 2, 27, 37, 45, 68, 69, 71, 72, 88, 100"
    },
    {
      "url": "https://huggingface.co/MeiGen-AI/MeiGen-MultiTalk/raw/b3ccbea2f68c89fafb277b9bd907905fff7a9337/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt lines 43, 45, 47, 48, 49, 52, 62, 86"
    },
    {
      "url": "https://huggingface.co/Soul-AILab/SoulX-FlashTalk-14B/raw/f3fda6499a6022862cadd39fc41ae63b64706e1b/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt lines 2, 16, 40, 76, 85, 103, 114, 118, 160"
    },
    {
      "url": "https://huggingface.co/Soul-AILab/SoulX-FlashHead-1_3B/raw/5fbd2e44ba08ab2deb28bd55cff2c15523e51ba5/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt lines 2, 11, 26, 27, 71, 109, 119, 124, 169, 173, 174, 175"
    }
  ]
}