Nougan Diffusers Loader 🚀

Most ComfyUI loaders expect a single .safetensors checkpoint. But a big chunk of the interesting 2026 models - Krea 2, a lot of HuggingFace-style releases - ship as a folder with unet/, text_encoder/, text_encoder_2/, and vae/ subdirectories, and ComfyUI's default loaders just stare at those and shrug. NouganDiffusersLoader is the node that reads the folder for you and hands you a proper MODEL, CLIP, and VAE, so you can wire it straight into the Krea 2 tooling that shares this pack.

How it works

Point it at a model folder and it hunts down the right files by conventional names: it looks for the UNet (diffusion_pytorch_model.safetensors or the fp16 variant), any text_encoder / text_encoder_2 dirs for CLIP, and the vae subfolder. It also accepts a single unified model file and loads that directly. Precision is auto-detected from the files - there's no weight-type dropdown to forget, which is one less thing to get wrong on Krea 2's fp8 setups.

The interesting part is the attention handling. Two widgets let you pick a SageAttention or FlashAttention version, and the loader applies the matching attention patch to the model when you run. It's wired to be compatible with both attention-patch calling conventions across ComfyUI versions (some pass pe/attn_mask as their own kwargs, some bundle them in extra_options), including a manual rotation-matrix RoPE fallback for Flux/Krea-family models if ComfyUI's own apply_rope isn't found. That's the kind of thing that looks like boilerplate until you load a model and the reshape doesn't explode.

The inputs that matter

• model_name - dropdown of everything found in your diffusion_models/, unet/, and diffusers/ roots. Pick folder or file.
• sageattention_version - None, SageAttention 2, or SageAttention 3. Requires pip install sageattention.
• flashattention_version - None, FlashAttention 2/3/4. Requires pip install flash-attn.
• model_data - a JSON blob managed by the node's UI. You don't touch it, and notably the loader now ignores it on purpose: an earlier bug let a stale UI populate it and clobber your saved model selection on reload, so the fix was to make the native widgets the single source of truth.

Outputs: MODEL, CLIP, VAE.

Installing

Search Nougan in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
# restart ComfyUI

The pack itself has zero pip dependencies. SageAttention and FlashAttention are optional - leave both on None and everything still works, just slower. If you do want them, install them into the same Python environment ComfyUI uses, and note that FlashAttention is famously version-picky about CUDA builds. SageAttention 2 is the more forgiving of the two on consumer cards.

Where people get burned

The two real failure modes are a model folder it can't find (Model 'X' not found in any scanned root - check which roots are being scanned in the boot log) and a folder with an oddly-named UNet file it can't match, which throws a "No UNet file found" error. Both are visible in the console prints, so the fix is usually a rename or a rescan. The attention dropdowns are the trap only if you install Sage/Flash and forget the matching Python package - the node will tell you the requirement in the tooltip, but it won't fall back silently for you.