Load Diffusion Model Hub Pro

Load Diffusion Model Hub Pro is the UNET-flavored sibling of the pack's flagship hub. Where Load Checkpoint Hub Pro loads a whole checkpoint and hands you MODEL, CLIP, and VAE, this one loads just the diffusion/UNET weights and routes the sampling recipe alongside: steps, refiner steps, CFG, sampler, scheduler, denoise, plus your positive and negative prompts and a metadata string. One node, the model and the entire generation plan, on a single line of sockets.

It's the node for the modern split-model world. Checkpoints bundle everything into one file; diffusion models don't. If you're running Flux, SD3, or any of the 2026 generation that ships its text encoder and VAE as separate downloads, you're already working with a diffusion_models folder and a grab-bag of companion files - and this hub matches that reality by loading only the UNET and leaving CLIP and VAE to you.

How it works

Under the hood it's comfy.sd.load_unet() on the resolved model path, wrapped in the same parameter-routing surface as the checkpoint hub. The unet_name dropdown is built from folder_paths.get_filename_list("diffusion_models"), so it covers both models/diffusion_models and models/unet. The sampler and scheduler lists come live from comfy.samplers, so they always match what your ComfyUI supports. Prompts are multiline with dynamic prompts enabled; metadata is single-line.

The outputs, and the one that's easy to miss

• MODEL - the loaded UNET weights, straight into your sampler.
• steps, step_refiner, cfg, sampler_name, scheduler, denoise - the recipe, wired to the matching KSampler inputs. The sampler and scheduler sockets are combo-typed, so they click into place.
• unet_name (*) - the filename as a wildcard, for routing into context nodes or loaders that want to know which model ran.
• positive_prompt, negative_prompt, metadata - plain STRING outputs feeding your CLIP Text Encode nodes and save/metadata nodes.

Here's the part people trip on: there is no CLIP and no VAE output. The node only loads the diffusion model. You still need a separate CLIP loader (or DualCLIP loader) and a VAE to actually encode your text and decode your latents - and in 2026 those are separate files you download yourself. This isn't a bug; it's the node doing the one job a diffusion_models folder actually contains. If your workflow errors on a missing CLIP or VAE, that's the gap.

Installing it

Part of MarwanDSAI/comfyui-mdsnodes. ComfyUI Manager: search "ComfyUI-MDSNodes", install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/MarwanDSAI/comfyui-mdsnodes

Restart, no pip step - requirements.txt is empty and the pack declares zero dependencies.

Gotchas

If your diffusion_models folder is empty, the dropdown falls back to a single "None" entry - and queuing with that selected will error out, because there's nothing to load. Drop at least one model in first. And remember the hub is explicit-wire plumbing, not a magic context bundle: the benefit is one authoritative source for the recipe, and the cost is that you're still wiring the sockets by hand. That's the trade, and for a testing workflow it's usually the right one.