VAE Decode (MultiGPU)

Let's get the naming straight, because the pack itself does the work of correcting it. VAE Decode (MultiGPU) does not shard the VAE across your GPUs. It's a thin wrapper around ComfyUI's own VAE decode - same operation as the stock VAE Decode node - that adds a status string and handles one edge case. The "MultiGPU" is branding, not mechanism. In this pack, only the UNet loaders actually distribute; the CLIP and VAE sit on the primary GPU.

So why would you reach for it? Three reasons, and they're honest ones. It keeps every stage of a multi-GPU graph in the same node family, so a workflow built with this pack reads consistently. It gives you a status output you can log - the pack's whole diagnostic story runs on status strings landing in comfyui.log. And it quietly handles 5D video-shaped latents, flattening the temporal/batch dimension down to a normal 4D image tensor instead of erroring. That last bit is genuinely useful if you ever pipe a video model's latent through this graph.

How it works

Inputs are samples (LATENT) and vae (VAE). It calls vae.decode directly, checks whether the result is 5D, and if so reshapes it to (batch, H, W, C). Outputs are image (IMAGE - wire into a Save Image or Preview node) and status (STRING - "✅ VAE decode finished", or an error message on failure).

One behavior worth knowing: if decode raises, it doesn't let the graph die - it returns a random-noise placeholder image plus a status describing the failure. That keeps the graph "working" so you can see the error, but it also means a silently failed decode hands you static instead of an exception. Read the status string.

Installing it

It ships with Comfy-MultiGPU-Loader. ComfyUI Manager search "Comfy-MultiGPU-Loader", or:

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

Restart, find it under MultiGPU/VAE.

The takeaway

If you already have a stock VAE Decode in a working workflow, there is no performance reason to swap it for this. Swap it if you want the status-string plumbing in the same graph, or if you're feeding it video-shaped latents. And if you came here hoping this node spreads a VAE across cards - it doesn't, and no node in this pack does. The project is sunset/"as-is" (GPL-3.0, no active support), and this node is the clearest example of the pack's honest scope: multi-GPU for the heavy part, plain wrappers for the rest.