what is comfyui-krea2-styletransfer fr?

This project adds a set of custom nodes to ComfyUI, a visual tool for building AI image generation workflows, that let you copy the visual style of one reference image onto a new image while keeping the content described by your text prompt. It works with the Krea2 image model and does not require any training or preparing a dataset, unlike the more common approach of training a LoRA, which is a small add-on file that teaches a model a particular look. The README explains that Krea2 does not officially ship a style reference feature, and the usual workarounds each have problems. Image-to-image editing tends to change the subject or layout instead of just the style, and extracting a style through text prompts alone is unreliable and depends heavily on how the description is worded. Earlier attempts at directly injecting a reference image's style into the generation process could transfer the look, but often let unwanted details from the reference image leak into the output, or reduced overall image quality. The author's main technical finding is that two settings, called low_scale_end and ref_k_strength, can be separated from each other. Normally, turning up style transfer strength also let content leak in and hurt quality, while turning it down protected quality but weakened the style transfer. By adding ref_k_strength as its own separate control, the project lets low_scale_end stay set low, which keeps content leakage and quality problems in check, while ref_k_strength brings the style transfer back on its own. The project includes several ComfyUI nodes: one to prepare a single reference image, one to run the main single-image style transfer, and a pair of nodes for an experimental two-reference mode that blends the style of two images together in stages rather than simply averaging them. The two-reference mode is deliberately capped at two images, since the README notes that three or more references tend to compete with each other and produce unstable or hard to control results. Recommended settings for both strength values and sampler parameters are provided directly in the README for users who want to start from a tested configuration.