Cloud LTXV Audio VAE Decode

LTX-2 is the video model that puts audio in the same latent space as the frames - you generate a clip and its soundtrack as one joint tensor. But a latent blob of audio isn't sound; something has to decode it into actual audio samples. That's this node. Cloud LTXV Audio VAE Decode takes your CLOUD_LATENT audio and a CLOUD_AUDIO_VAE, contributes an LTXVAudioVAEDecode step to the cloud workflow JSON, and hands back a CLOUD_AUDIO handle.

Now, about that handle type: CLOUD_AUDIO is not audio you can play yet. The pack's handles are all lightweight references that carry accumulated workflow JSON, and this pack is deliberately split into "cloud-side" nodes (build the JSON) and terminal "bridges" (fetch the result as a real tensor). CLOUD_AUDIO is cloud-side. To get an actual AUDIO tensor in your graph you wire it into Cloud Fetch Audio, the node that downloads the result and hands you something you can preview or pass to a local audio node.

Inputs

Only two, and both are just routing:

• samples - a CLOUD_LATENT. In an LTX-2 AV workflow this is the audio half you got from Cloud LTXV Separate AV Latent after sampling, or the output of Cloud LTXV Empty Latent Audio if you're just probing.
• audio_vae - a CLOUD_AUDIO_VAE handle from Cloud LTXV Audio VAE Loader. It's a distinct handle type on purpose: you can't accidentally wire the image VAE in here.

Output is one Audio (CLOUD_AUDIO) ref.

Installation

Part of the ComfyUI-CloudAPI-worker pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker

Copy config.json.example to config.json, paste an API key from https://platform.comfy.org/profile/api-keys, restart. In ComfyUI Manager by pack name; deps are only requests, Pillow, safetensors.

Gotchas

• It is not the terminal node. If you wire CLOUD_AUDIO straight into a preview node you'll get a type mismatch. Chain: ... sampler → Cloud LTXV Separate AV Latent → this node → Cloud Fetch Audio → real audio.
• Remember this whole pack is a proof of concept, and the LTX-2 audio path is its newest, least-tested corner. The included example_LTX_workflow.json is the reference for how the AV latent chain is supposed to connect; when in doubt, mirror that graph with the Cloud versions.
• Long jobs: the terminal bridge that actually fetches audio is where you bump the timeout for big clips, not this node.