Krea2 Model Loader ⚡

ComfyUI already knows Krea2 natively - the architecture is detected by core, CLIPType.KREA2 selects its text encoder, there's no secret sampling format. What core doesn't give you is a GGUF-aware loader for it. Krea2 Model Loader is the thin convenience wrapper that fills that gap: MODEL/CLIP/VAE by name, with GGUF quants staying quantized through the pack's custom ops. It's the boring, reliable front door to every Krea2 workflow, and boring is exactly what you want from a loader.

What it loads

Three dropdowns, three standard folders:

• unet_name - the Krea2 diffusion model from models/diffusion_models (unet). .safetensors or a GGUF quant. Krea2 is 12B - bf16 is ~24.8 GB - so this is where the quant decision happens. FP8 and GGUF Q4-Q8 are the normal choices on consumer cards.
• clip_name - the Qwen3-VL-4B text encoder from models/text_encoders. People routinely miss this one and the VAE, then report "Krea2 is broken". It's not broken; there are three files.
• vae_name - Krea2's VAE (the Qwen-Image VAE, ~250MB) from models/vae.

Outputs: model, clip, vae - standard comfy objects. From there it's Krea2Img2Img → Krea2 KSampler (or stock KSampler) → VAE Decode.

Why a separate loader at all

Krea2 is natively detected by core, so technically you could assemble the three with stock nodes - UNETLoader plus CLIPLoader (type: krea2) plus VAELoader. This node collapses that into one dropdown set and, more importantly, routes the GGUF files through this pack's GGUF loader so a quantized Krea2 loads without you hand-wiring a GGUF UNET loader and a GGUF CLIP loader next to each other. It matches the shape of the pack's other convenience loaders (ZImageLoader, QwenImageModelLoader, FluxKleinModelLoader), so if you know one you know them all.

Installing it

Same pack, one install:

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart; it's under 🤖 CCTech/Krea2. The usual pack-wide note applies: keep ComfyUI recent enough for the custom ops, and don't skip the gguf pip upgrade - the "No module named gguf" import error is the classic first-run failure.

Gotchas

• Empty dropdown = missing file. GGUF Krea2 files go in models/diffusion_models, the encoder in models/text_encoders, VAE in models/vae.
• Don't grab an mmproj or a random Qwen3 for clip_name - Krea2 expects the Qwen3-VL-4B encoder specifically. This pack's CCTechClipProjLoader and its Qwen3-TTS nodes use different encoders for different jobs; don't cross them.
• Licensing: Krea2 ships under its own community license - worth reading before you put outputs anywhere commercial. The model's fine to run; the license is on you.

The loader does exactly one job and stops. Every interesting Krea2 thing happens in the img2img node and the Control LoRA loader downstream of it.