extraltodeus/comfyui-krea2-attention-tweak

This is a very small custom node for ComfyUI, the node based interface many people use to run Stable Diffusion style image generation workflows. It patches how the attention mechanism behaves inside the Krea2 Turbo image model, specifically the part of the model that decides how strongly it should follow your text prompt versus generating more freely on its own. Attention, in this context, is the internal mechanism a model uses to decide which parts of your prompt matter most when producing an image, and adjusting it changes the balance between obeying your wording exactly and taking more creative liberty. The author based this patch on a separate LoRA file made by another creator, called Krea-2-Turbo-Projector-Scale-LoRA-Diffusers, hosted on Hugging Face. A LoRA is a small add on file that nudges an existing image model's behavior without retraining the whole thing. While comparing that LoRA against the original model, the author noticed its internal values looked close enough to the unmodified model to work as a simple multiplier rather than a full retrained adjustment. This patch turns that observation into a small, adjustable setting you can apply directly inside ComfyUI, rather than requiring you to load the separate LoRA file at all. The README states that setting this multiplier to 3.0 produced results the author found more balanced, sitting between stronger creativity and closer prompt following. No other settings, ranges, or recommended values are given beyond that one example, and no comparison images are included to show the effect visually. The README is very short and does not include installation steps, a license, or example images. There is no description of how the node is added to a ComfyUI workflow graph or what its input and output connections look like. This should be treated as a small, experimental tweak for people already comfortable installing and using custom ComfyUI nodes, rather than a polished or fully documented tool.