Load Diffusion Model X

If you test workflows across a few checkpoints - a Flux dev here, a Qwen image there, an SDXL finetune when a workflow demands it - you're used to dragging the wrong model into a loader and hunting through the dropdown every time. Load Diffusion Model X lets you register your preferred diffusion models as named rows in one node and switch the active one with a single click. The node only loads the active row, so you're not loading everything at once - just keeping your alternatives visible on the canvas.

Under the hood it's deliberately boring in a good way: the active row is loaded through ComfyUI's native UNETLoader, and the shared weight_dtype dropdown (default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) applies to whichever row is currently active. It reads the same weight-dtype choices and behavior as the native loader, so nothing surprising happens at the model level - the node is a picker organizer, not a reimplementation. One MODEL output carries whatever the active row produced.

The part that's genuinely useful

The row management is where the effort went: add, replace, remove, search, refresh, inspect model details, and radio-style active toggles. The picker pulls metadata and previews from LoRA Manager when it's installed, but it stays fully usable from ComfyUI's plain diffusion-model folders if it isn't. That means the node scales from "I have two checkpoints" to "I maintain a rotation of ten," without the dropdown getting any harder to use.

Why you'd reach for it: you're building configurable workflows where the checkpoint is a switchable input, or you're tired of remembering which folder a model lives in. Pair it with WorkflowX's config system and different active rows can even resolve per configuration - same workflow, different checkpoint per profile.

Inputs:

• weight_dtype - the shared precision dropdown for the active model.

Output:

• MODEL - the loaded active model.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/haroonaslam/WorkflowX-Configurator

or search WorkflowX Configurator in ComfyUI Manager, restart ComfyUI, refresh the browser. No pip dependencies; the models themselves live in your normal models/diffusion_models folder.

The gotchas

• It loads the active row only. Switching active rows on the canvas triggers a reload at run time - that's the point, but on a slow disk it's not instant.
• weight_dtype applies to all rows. If you want different precision per model, this single shared dropdown can't express it. It's one knob for the whole node.
• Your model files still have to be there. The node keeps the list visible but doesn't download anything - a missing file fails the same way the native loader would.
• fp8 isn't magic. fp8_e4m3fn_fast trades a hair of quality for speed; if a model looks subtly off, try default before assuming the checkpoint is bad.

A quality-of-life loader that shines for people who rotate models. If you mostly run one checkpoint forever, the native loader is fine - this earns its keep when the active model is a variable, not a constant.