Icy DiffusersLoader
When your checkpoint is a folder, not a file
- MODEL
- CLIP
- VAE
IcyDiffusersLoader is the ComfyUI-IcyHider wrapper around core DiffusersLoader - same node, renamed class so the preview-hiding extension can cover it. And before we go further: if you're a typical ComfyUI user, you probably never need this node. But when you do, it's the only thing that works, so it's worth understanding.
The node loads a complete diffusion model from a diffusers-format directory instead of a single checkpoint file. Where a normal checkpoint is one big .safetensors, a diffusers model is a folder tree - separate subdirectories for the UNet/transformer, the text encoder(s), the VAE, each with its own config and weights. It's the format HuggingFace's own library (diffusers) uses natively, and it's what you get if you pull a repo with git clone https://huggingface.co/... rather than downloading the merged single file. The KB's file-formats material notes that most distributions have converged on single-file .safetensors because it's simpler and safer; the folder format survives mostly for training checkpoints and people who prefer raw diffusers structure.
Inputs and outputs
- model_path (enum) - a directory under
ComfyUI/models/diffusers/. ComfyUI scans that folder at startup, so a directory you drop in needs a restart to appear. - Outputs: MODEL, CLIP, VAE - the three components separately, exactly like
CheckpointLoaderSimplebut as three distinct sockets ready to be wired into your sampler, text encoder, and VAE decode.
The practical upside of the split output: you can swap just the VAE or just the text encoder on a diffusers model without loading the whole thing again. The downside is verbosity - three wires where a normal loader gives you one.
When you'd actually use it
- You downloaded a model that only exists as a diffusers folder (some newer or experimental HuggingFace repos ship this way).
- You're doing training or testing with a model you've exported in diffusers format.
For everyday generation, CheckpointLoaderSimple (or a GGUF loader for quantized Flux) is almost always the right call. This node is the fallback for the format minority.
Installing it
cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider
Restart, or install via ComfyUI Manager (search "IcyHider"). No pip dependencies, no models bundled - you supply the diffusers directory.
Common issues
- Missing node in shared workflows. Install the pack, or substitute core
DiffusersLoader. - Folder not in the dropdown. It has to live in
ComfyUI/models/diffusers/and ComfyUI has to have been restarted since you put it there. - Load errors on a folder that looks right. Diffusers checkpoints come in many sub-formats (SD 1.5, SDXL, Flux, SD3 all have different directory layouts). If the loader rejects yours, it's usually a format/architecture mismatch, not a broken install.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| model_path | COMBO | 0 options: |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| CLIP | CLIP | — |
| VAE | VAE | — |