Krea-2 in ComfyUI: setup and reference image tutorial

Nerdy Rodent explains how to build Krea-2 workflows in ComfyUI for local AI image generation. The tutorial covers NVFP4, FP8 and BF16 options. The speaker uses the 8 GB NVFP4 download for most examples and finds little visual difference between the model variants in the prompt comparison shown. That download size is not a stated VRAM requirement. Eight sampling steps at CFG 1 are the suggested starting settings.

Reference images can guide style and composition through image conditioning, with up to three inputs. The speaker stresses that Krea-2 is not an editing model: references act as suggestions. For image-to-image generation, he generally recommends denoise below 0.5 to retain some of the original, depending on the sampler and scheduler. A three-reference example uses 0.3 denoise with the linear quadratic scheduler to address images appearing as a grid.

Other examples cover style changes, aspect ratio changes and inpainting with a masked latent. SAM 3 supplies the mask in the inpainting demonstration. Upscaling options include refinement passes and a 2× VAE, which the speaker says can take 2048 × 2048 output to 4096 × 4096. Conditioning rebalancing may improve prompt adherence, but he reports that it can also reduce text clarity. He also cautions that Krea-2 is not open source and advises checking its license, especially for corporate use.