VAE Decode (MultiGPU Debug)

There's a pattern in this pack you'll spot by the time you reach the debug nodes: each one is a subclass of its production twin, re-registered under MultiGPU/Debug, with identical inputs and outputs. VAEDecode (MultiGPU Debug) is the most honest example, because the parent VAEDecodeMultiGPU is already just a thin wrapper around ComfyUI's stock VAE decode - so this debug variant adds exactly nothing functionally. It's the same code path, the same samples (LATENT) + vae (VAE) inputs, the same image + status outputs.

What it is for: consistency inside a debugging graph. The pack's USER_GUIDE workflow is a chain of debug nodes - debug loader, debug sampler, display - and if you're logging statuses to comfyui.log, you want the decode stage reporting in the same style, in the same menu. That's the entire value proposition, and it's a modest one.

Like its parent, it calls vae.decode, flattens any 5D temporal latent to a normal 4D image, and returns a status string. On a decode failure it returns a random placeholder image rather than killing the graph - which is a trap if you don't read the status, so read it. There's no multi-GPU mechanism in this node; the VAE lives on the primary GPU in this pack's architecture. Only the UNet loaders actually shard.

Installing it

Ships with Comfy-MultiGPU-Loader:

cd ComfyUI/custom_nodes
git clone https://github.com/AngelCookiesLab/Comfy-MultiGPU-Loader
pip install -r requirements.txt

or ComfyUI Manager search "Comfy-MultiGPU-Loader", then restart. It appears under MultiGPU/Debug.

Bottom line

Use this node in a debug graph because it keeps the family together; use the plain VAEDecodeMultiGPU or stock VAE Decode anywhere else. It's a placeholder in the truest sense - the debug menu slot for a stage that has nothing to debug. Remember the project status: sunset, GPL-3.0, as-is, no active support. This is the pack's least interesting node, and that's fine - not every socket needs a story.