Icy UNETLoader

IcyUNETLoader is the IcyHider clone of UNETLoader, the node you use when you want to load the diffusion model itself - the denoiser - without the CLIP text encoder and VAE bundled around it. In 2026 you mostly hit this with two kinds of files: models released as separate UNet/diffusion components, and quantized GGUF or fp8 variants you want to pair with a specific text encoder and VAE of your own choosing.

What it does

One required input that decides everything:

• unet_name - the model file, picked from a dropdown populated by models/diffusion_models (the folder UNETLoader reads). This is the folder that trips people up: drop the file in the wrong place and the dropdown stays empty no matter how many times you refresh.

One input that matters when the file supports it:

• weight_dtype - how the weights are loaded: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2. fp8 variants trade a little quality for a large chunk of VRAM - on a card where the model barely fits, fp8 is often the difference between running and not. fp8_e4m3fn_fast is the go-to when speed matters more than the last drop of precision. If the file isn't fp8-quantized, leave default alone.

Output is a MODEL - the denoiser - which you wire into your sampler stack.

How it fits

The pattern is "split checkpoint": load the denoiser with UNETLoader, the text encoder with a CLIPLoader (or the checkpoint's CLIP), and the VAE with a VAELoader, then build the graph from those three pieces. It's the standard structure for Flux-style setups and for quantized models where the checkpoint file doesn't carry everything. People also reach for it to mix - this node's UNet with a different VAE or encoder than the release intended.

Where people get burned: loading a regular full checkpoint (.safetensors with everything baked in) through UNETLoader and getting errors - that file belongs in models/checkpoints, not models/diffusion_models. Also, on fp8 files, running the wrong weight_dtype can produce garbage or a crash; match it to what the model page says it is.

The "Icy" caveat

Outputs MODEL, no image preview, so the Icy wrapper is cosmetic packaging - the same class ComfyUI calls UNETLoader, relabeled by the pack's automatic clone-everything mechanism. Use the core node if that's what your workflow template references; they're identical.

Install

ComfyUI Manager → search "IcyHider" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider.git

Restart ComfyUI. No Python dependencies and no models bundled - the diffusion model is yours to source. The node lives in IcyHider Comfy Core. Empty unet_name dropdown? File's not in models/diffusion_models, or you skipped the restart.