wtf is cross-architecture-weight-grafting?

This project is an experimental technique for combining parts of two different AI image generation models even when they were built with different architectures. Normally, merging models only works if both share the same structure and layer sizes, so models like Qwen Image, Krea, FLUX and Klein cannot be merged with standard tools. The author calls this approach cross architecture weight grafting: instead of merging entire models, it inspects the layers of both models, matches up blocks that seem to do similar jobs, resizes tensors where needed, and blends in only a small percentage of one model's weights into another. The goal is not to replace either model, but to let one borrow some visual qualities from another while mostly keeping its own behavior. The best result documented so far grafts a small amount of Krea's weights into Qwen Image, keeping Qwen's strong composition and prompt understanding while picking up some of Krea's lighting and texture quality. The author found that keeping the blend ratio low, only a few percent per layer, works best, and that pushing it higher usually damages image quality instead of improving it. Other combinations, such as merging in the opposite direction or grafting between FLUX and other architectures, produced weaker or inconsistent results. The repository does not include any finished merged model files, since the original models carry different licenses that restrict sharing modified weights. Instead it provides the ComfyUI workflow, configuration files, and example images so others can reproduce the experiments themselves using models they download from the original sources. It runs on CPU only, and the amount of system memory needed depends mainly on the size of the target model rather than the donor model. This is explicitly framed as an ongoing research experiment rather than a finished, ready to use tool.