ComfyUI Node
K2 Load Krea 2
Convenience loader for the three Krea 2 components. Equivalent to UNETLoader + CLIPLoader(type=krea2) + VAELoader, just tidier. Use the native loaders instead when the graph needs quantization, caching or device nodes in between — everything downstream works with either. The text encoder must be a Krea-compatible Qwen3-VL-4B; a generic encoder produces the wrong 12-layer conditioning layout and Krea will refuse it.
K2 Load Krea 2
- model
- clip
- vae
◄diffusion_model▾►
◄text_encoder▾►
◄vae▾►
◄weight_dtypedefault►
◄text_encoder_devicedefault►
CategoryFVM Tools/K2
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| diffusion_model | COMBO | Krea 2 Turbo or RAW transformer. | |
| text_encoder | COMBO | Qwen3-VL-4B encoder for Krea 2. | |
| vae | COMBO | Qwen-Image VAE (Krea 2 shares its autoencoder). | |
| weight_dtype | COMBO | default | 'default' follows the file. FP8 modes lower model memory where the device supports them; fp8_e4m3fn_fast also enables ComfyUI FP8 matmul. |
| text_encoder_device | COMBO | default | 'cpu' keeps Qwen off the GPU — slower prompt encoding, noticeably less VRAM. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| clip | CLIP | — |
| vae | VAE | — |