Nodes/ZONOS2 TTS/ZONOS2 Model Loader
ComfyUI Node

ZONOS2 Model Loader

Load native ZONOS2 weights with ComfyUI/AIMDO memory tracking.

By Saganaki22·Created 2 months ago·Updated 2 months ago· 12
ZONOS2 Model Loader
    • zonos2_model
    model
    dtypeauto
    attentionauto
    download_if_missingtrue
    CategoryZONOS2 TTS

    Inputs (4)

    NameTypeDefaultDescription
    modelCOMBOCheckpoint in ComfyUI/models/zonos2. The BF16 and mixed FP8 presets download from their own Hugging Face repositories. Mixed FP8 is detected from metadata and uses FP8 only for MoE expert gate/up weights.
    dtypeCOMBOautoRuntime weight dtype. Auto preserves a standard checkpoint's dtype and uses BF16 compute for mixed FP8. Mixed FP8 accepts auto or bf16, not fp16. Changing this setting hard-unloads the previous bundle.
    attentionCOMBOautoAttention backend. Auto selects FlashAttention on compatible CUDA BF16/FP16 systems and otherwise uses PyTorch SDPA. Changing this setting hard-unloads the previous bundle.
    download_if_missingBOOLEANtrueWhen enabled, independently download only missing assets from the selected preset's Hugging Face repository. An existing DAC or speaker encoder is reused, while a missing selected checkpoint is still downloaded. When disabled, every required asset must already exist under ComfyUI/models/zonos2.

    Outputs (1)

    NameTypeDescription
    zonos2_modelZONOS2_MODELLoaded ZONOS2 model bundle containing the native language model and DAC decoder, managed by ComfyUI/AIMDO.