HuMo 17B and 1.7B: human video from text, reference images and audio (TA and TIA modes, Wan2.1 and Whisper dependencies, 97-frame training length)

HuMo: human video from text, a reference image and audio What it is. HuMo, from Tsinghua University and ByteDance's Intelligent Creation Team (paper arXiv 2509.08519), is a human-centric video model conditioned on text, reference images and audio. The authors describe strong prompt following, consistent subject preservation and audio-synchronized motion. Those are their claims, not results checked here. Which input modes you can actually run. The card advertises three uses: text plus reference images (to customize appearance, clothing, makeup, props and scenes), text plus audio, and text plus image plus audio. At the pinned revision, its own to-do list marks Text-Audio and Text-Image-Audio inference as released but Text-Image inference as not yet released, and the config documents only mode "TA" (text+audio) and "TIA" (text+image+audio). So for reference-image identity work, the documented route also needs an audio track; a reference-image-only run is not documented here. Checkpoints and dependencies. HuMo-17B supports 480P and 720P. HuMo-1.7B is listed as "Lightweight on 32G GPU"; the authors say it generates a 480P video in 8 minutes on a 32G GPU, with lower visual quality than 17B but nearly unaffected audio-visual sync. No memory figure is given for 17B. HuMo also needs Wan-AI/Wan2.1-T2V-1.3B (for its VAE and text encoder), openai/whisper-large-v3 (audio encoder) and, optionally, huangjackson/Kim_Vocal_2 to remove background noise from the audio. The card's single download command for bytedance-research/HuMo names no subfolder, while the card links separate HuMo-17B and HuMo-1.7B folders in that repository. Setup as documented. Python 3.11; torch 2.5.1, torchvision 0.20.1 and torchaudio 2.5.1 from the CUDA 12.4 wheel index; flash_attn 2.6.3; pip install -r requirements.txt; ffmpeg from conda-forge. Prepare the text, reference images and audio as in examples/test_case.json. Run bash scripts/infer_ta.sh or scripts/infer_tia.sh for 17B, and scripts/infer_ta_1_7B.sh or scripts/infer_tia_1_7B.sh for 1.7B. Multi-GPU inference uses FSDP plus sequence parallelism; set dit.sp_size to the number of GPUs. Settings in humo/configs/inference/generate.yaml. generation.frames sets the length. scale_a is the audio guidance strength (higher gives better audio-motion sync, per the comment) and scale_t is the text guidance strength (higher gives closer prompt adherence). mode is "TA" or "TIA". Height is 720 or 480 with width 1280 or 832. The shown sampling value is 50 steps; the comment says 30 to 40 is faster. The card gives no default values for frames, scale_a or scale_t. Length and resolution. The model was trained on 97-frame clips at 25 FPS, which is about 3.9 seconds by arithmetic. The authors warn that generating more than 97 frames may degrade results and say a longer-generation checkpoint will follow. They state that 720P gives much better quality than 480P. ComfyUI. The authors report (13 September 2025) that the 17B model is merged into kijai's ComfyUI-WanVideoWrapper. That integration was not checked here. Gaps. This card links no code repository: requirements.txt, the scripts, examples/test_case.json and generate.yaml are repository-relative paths whose host was not established here, and the Hugging Face file list was not checked. The number of reference images, per-mode defaults, output frame rate and 17B memory needs are not stated. A Best-Practice Guide is announced but not released at this revision. Rights. The card metadata declares Apache-2.0. Wan2.1, Whisper and the audio separator carry their own terms, which were not reviewed. The card states no consent or likeness policy, and this record grants no likeness or consent clearance; obtain permission for any real person you depict. Based on the bytedance-research/HuMo card pinned to revision 3a4a1610d399a5cbb932d54dc229944029803ff7 and observed on 2026-10-28. Nothing was downloaded or run. Quality, subject preservation, audio sync and the 1.7B timing are the authors' statements; nothing was run here. Repository-relative files (requirements.txt, scripts, examples/test_case.json, generate.yaml) and the reported ComfyUI integration were not inspected. The card's licence label is a declaration; terms for Wan2.1, Whisper-large-v3 and Kim_Vocal_2 were not reviewed. 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; UTF-8 bytes 2562:2868, 2917:3383, 3504:3722, 1904:2539, 2093:2539, 3767:4035, 4698:5076, 5042:5076, 5517:6005, 5831:6005, 6586:6812, 7098:7205, 4:138; exact excerpt SHA256 3e8424b336c86d60228ed2dec5465d03b2f6e88c27342728685f5aae1fb5b68c,c6d732d67cad699f580d34aef671d444301c8ff746ed6a46a37f9"
    }
  ]
}