Nodes/ComfyUI_StarNodes/⭐ Star VAE LTXV Load
ComfyUI Node

⭐ Star VAE LTXV Load

Star VAE LTXV Load — two VAEs, one LTX video node

By Starnodes2024·Created 2 years ago·Updated 2 days ago· 106
⭐ Star VAE LTXV Load
    • video_vae
    • audio_vae
    video_vae
    audio_vae

    LTX Video's newer releases are the odd one out in the video-model world: they don't have a single "VAE" - they ship two, one for video frames and one for audio. If you've tried to wire LTX video into a stock ComfyUI graph, you've hit the wall where there's no obvious way to load both. Star VAE LTXV Load exists for exactly that gap: it loads the video VAE and the audio VAE side by side from your models/vae folder and hands you two proper VAE sockets.

    Why two VAEs?

    LTXV (and especially the LTX-2 line from Lightricks) is built around a combined model that generates video and its soundtrack. The video side uses a standard spatial-temporal VAE to go between pixels and latents; the audio side uses a separate lightweight audio VAE that compresses waveforms into the latent space the model conditions on. The "all-in-one" LTX nodes in this same pack load both internally, but when you're hand-building a graph you need them as explicit objects - which is what this node provides.

    How it works

    Two dropdowns, two outputs:

    • video_vae - pick the video VAE file from models/vae. It's loaded as a standard ComfyUI VAE (comfy.sd.VAE), so it plugs into VAEEncode/VAEDecode like any other.
    • audio_vae - pick the audio VAE file. This one loads through the dedicated Lightricks AudioVAE class, because a stock VAE loader would have no idea what to do with an audio codec.

    Outputs are video_vae (VAE) and audio_vae (VAE), matching the input order. Wire the video one into your encode/decode chain and the audio one into whatever LTX audio node you're using.

    Where the files come from

    The one real gotcha: models/vae has to contain both files, and LTX checkpoints ship with them inside the checkpoint rather than as loose files. That's what the companion Star VAE LTXV Save node is for - run it once on your LTX checkpoint and it splits the video and audio VAEs out into models/vae as separate .safetensors. Alternatively, Lightricks' own LTX downloads page provides the split VAE files directly. Until one of those happens, the dropdowns will be empty.

    Installing

    Part of the StarNodes pack:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
    cd ComfyUI_StarNodes
    pip install -r requirements.txt
    

    Restart, search star under ⭐StarNodes/Video. ComfyUI Manager users: search "Starnodes".

    Notes

    If you're running the pack's own Star LTXV All-in-One node, you don't need this - that node loads everything internally. Reach for Star VAE LTXV Load when you're building a manual LTX graph or a custom video+audio pipeline and you want the two VAEs as first-class sockets. Load both at once, and remember the audio VAE only makes sense in an LTX audio context; you can't decode a music file with it and expect a waveform you can audition directly.

    Category⭐StarNodes/Video

    Inputs (2)

    NameTypeDefaultDescription
    video_vaeCOMBO0 options:
    audio_vaeCOMBO0 options:

    Outputs (2)

    NameTypeDescription
    video_vaeVAE
    audio_vaeVAE