DiffusersLoaderMultiGPU

Most checkpoints you'll ever load in ComfyUI are a single .safetensors file. DiffusersLoaderMultiGPU is for the other format you'll occasionally run into: a HuggingFace Diffusers pipeline - a whole folder, with subdirectories for unet/, text_encoder/, vae/, and scheduler config, instead of one packed file. It's the device-aware twin of core ComfyUI's DiffusersLoader: point model_path at that folder instead of a filename, and add a device to say where it lives.

Why you'd still hit this format

safetensors won the single-file format war years ago, for good reason - no pickle execution risk, fast mmap loading, one clean file to move around. But some models still get published first (or only) as a raw Diffusers repo, before anyone gets around to converting it: research releases, older Stable Diffusion checkpoints, or a repo you cloned straight from HuggingFace with no community single-file build yet. This node saves you the manual conversion step - load the folder as-is.

The inputs and outputs that matter

• model_path - the Diffusers-format folder itself, discovered from your models directory the way ComfyUI usually discovers checkpoints.
• device (default cpu) - pins the whole loaded pipeline there, or to a second GPU if you have one.
• Outputs: MODEL, CLIP, VAE - all three, because a Diffusers pipeline bundles the same three components a checkpoint does, just split across subfolders instead of packed into one file's tensor prefixes.

How to install it

ComfyUI Manager: search ComfyUI-MultiGPU, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/pollockjj/ComfyUI-MultiGPU

then restart. No extra Python dependencies for this node specifically - though if the Diffusers repo you're loading needs a custom pipeline class ComfyUI doesn't already support, that's a limitation of core ComfyUI's Diffusers support, not something this node changes.

Common issues & troubleshooting

Empty model_path dropdown. This means you haven't actually placed a Diffusers-format folder in the expected models location - a bare .safetensors filename won't show up here, and a folder in the wrong layout won't either. Confirm the directory has the subfolder structure (unet/, text_encoder/, vae/) a real Diffusers export has.

Whole-pipeline pin, no splitting. Like every other plain MultiGPU loader, this one moves the entire thing to one device - nothing partial. If the pipeline doesn't fit anywhere you have available, that's what DiffusersLoaderDisTorch2MultiGPU is for: it splits the diffusion weights across devices instead of pinning them whole.

You reach for this less often than you'd think. Given how thoroughly safetensors displaced the Diffusers-folder format for distribution, most models you actually want to run have a proper single-file (or split UNet+CLIP+VAE) release available - check for one before assuming you need this node. Keep it in your back pocket for the odd release that only ships this way.