Load Diffusion Model (with Name Clean)

When you're running FLUX, Z-Image, or any of the split-file workflows, you load the diffusion model (the UNet) on its own - separate from the text encoder and VAE. Load Diffusion Model (with Name Clean) is that loader, plus two upgrades over the stock UNETLoader: it searches both models/checkpoints and models/diffusion_models so you can stop caring where the file lives, and it hands you a cleaned filename string for auto-folder output naming.

It's the sibling of this pack's Load Checkpoint (with Name Clean) - same cleaning logic, same folder trick - but UNet-only, and with the one extra feature that makes it genuinely useful on its own: built-in fp8 loading options.

The inputs

• model_name - the combined, sorted dropdown of both model folders. The only place you choose the model.
• weight_dtype - this is the knob that earns the node its place. Options:
  • default - load as-is (or at whatever precision the file stores).
  • fp8_e4m3fn - load in fp8 with the common E4M3 float format. Roughly half the VRAM of fp16 for the same model.
  • fp8_e4m3fn_fast - same fp8, but with ComfyUI's fp8 optimizations enabled, trading a little quality/latency stability for speed.
  • fp8_e5m2 - the alternate E5M2 format, mostly seen on older quantized files.

If you've ever had to chain an extra "model to fp8" or dtype-conversion node into your UNet load, this is the same result with one dropdown instead of two nodes. Q8 is basically fp16 at half the size; when your model fits in fp8 comfortably, this is the easy button.

The outputs

• MODEL - the loaded diffusion model, wired into your sampler's model input exactly like any UNet loader output.
• model_name_clean - the stripped filename string (extensions, fp16/fp8/bf16/scaled-style markers, trailing dashes/underscores all removed). Append a / and use it as a Save Image filename_prefix to get per-model output folders.

Install

From the ComfyUI Manager search Easygoing, or:

cd ComfyUI/custom_nodes
git clone https://github.com/easygoing0114/ComfyUI-easygoing-nodes.git

Restart ComfyUI. No extra pip requirements; the pack needs a ComfyUI build with the V3 node API, so update ComfyUI if none of its nodes register.

Notes and gotchas

A few practical things. fp8_e4m3fn_fast is a speed/memory trade, not a straight upgrade - if you see quality or instability regressions on a particular model, drop back to the plain fp8_e4m3fn and compare. And keep in mind this loads only the diffusion model; you still need a separate CLIP/text-encoder loader and a VAE for a full generation. The clean-name folder trick works exactly like the checkpoint version - one loader, one filename_prefix with a trailing /, and your outputs start sorting themselves by model. That's the whole pitch, and for people juggling a handful of split-file checkpoints it removes a real daily annoyance.