Diffusion Model Enum Switch 10X (Soze)

The checkpoint loader is a bundle deal - MODEL, CLIP and VAE together. But a lot of modern workflows don't work that way. Flux, Hunyuan and the single-file diffusion-model ecosystem want you to load the UNet or DiT on its own, and for that ComfyUI has DiffusionModelLoader, which picks one file from models/diffusion_models. Diffusion Model Enum Switch 10X is that loader with a string key bolted on: ten diffusion-model slots, an enum_value decides which one actually loads, and only that one ever touches VRAM.

It's the diffusion-model sibling of the pack's Checkpoint Enum Switch, sharing the same lookup logic, the same ten-case shape, and the same "only the match loads" memory win.

How it works

Same matching core as the rest of the family: enum_value is compared against up to ten compare_text_N labels - case-insensitive and whitespace-trimmed by default, blank labels skipped, first match wins - and the model in the winning slot is the one loaded. default_model covers the no-match case. Blank dropdown = slot unused, so an empty slot can't hijack a run.

Two extras make it worth using over a bare loader:

• weight_dtype - cast the weights on load, with default, fp8_e4m3fn, fp8_e4m3fn_fast and fp8_e5m2 options. If you're running a big DiT on limited VRAM, fp8 is where you'll live; fp8_e4m3fn_fast trades a touch of precision for speed. Leave it on default if you don't know what you're doing and you're fine.
• allow_none - enabled, an enum_value of "none" skips loading entirely and the node outputs nothing for that run.

Outputs are leaner than the checkpoint variant: MODEL, plus model_name (what actually loaded), matched_index and status. No CLIP or VAE here - diffusion models don't bundle them; you'll grab those from a checkpoint or a separate loader.

When it's worth it

This is the node that makes CSV-driven Flux/Hunyuan workflows work: your CSV row names the diffusion model, and this node loads it. Or run A/B comparisons across three different model files keyed by a text input. If you never change models mid-workflow, plain DiffusionModelLoader is simpler - the enum version pays off exactly when a string, not a human hand, does the choosing.

Installing it

Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/SozeInc/ComfyUI_Soze.git
pip install -r ComfyUI_Soze/requirements.txt

Or via ComfyUI Manager ("ComfyUI_Soze") and restart. It reads whatever's in models/diffusion_models - no extra downloads. As with the checkpoint variant, there's a "Copy Names to Compare Text" button on the node that fills the case labels from your dropdowns and copies a JSON array for the enum_value source. Watch one thing: fp8_e4m3fn_fast wants a fast path in your ComfyUI install - if you get a loading error with it, fall back to fp8_e4m3fn rather than fighting it.