Nodes/Higgs v3 TTS/Higgs v3 Load Model
ComfyUI Node

Higgs v3 Load Model

Load Higgs Audio v3 TTS natively with ComfyUI/AIMDO memory registration.

By Saganaki22·Created 2 months ago·Updated 2 months ago· 64
Higgs v3 Load Model
    • higgs_model
    modelHiggs Audio v3 TTS 4B - bosonai (auto-download)
    dtypeauto
    deviceauto
    attentionauto
    download_if_missingtrue
    CategoryHiggs v3 TTS

    Inputs (5)

    NameTypeDefaultDescription
    modelCOMBOHiggs Audio v3 TTS 4B - bosonai (auto-download)Model folder under ComfyUI/models/higgsv3tts. Put model.safetensors in higgs-audio-v3-tts-4b or the root higgsv3tts folder.
    dtypeCOMBOautoWeight dtype for Higgs and its audio codec. auto uses bf16 on supported CUDA and fp32 otherwise. fp16 is hidden because it can produce non-finite audio.
    deviceCOMBOautoDevice for native inference. auto follows ComfyUI's current torch device; cuda is fastest; cpu is fallback only and very slow.
    attentionCOMBOautoAttention backend. auto/sdpa are usually fastest here; flash_attention needs flash_attn; sageattention may be slower for token-by-token TTS.
    download_if_missingBOOLEANtrueIf files are missing, downloads small assets plus the large model.safetensors into ComfyUI/models/higgsv3tts/higgs-audio-v3-tts-4b.

    Outputs (1)

    NameTypeDescription
    higgs_modelHIGGSV3TTS_MODEL