>>106313113
generally speaking yes, it can and will happen quite easily on a larger SDXL training, I'm thinking it's not just the TE either. the overall better performance of newer models like wan/chroma/qwen/... is also because of this.

will it happen so easily as a "hug" concept training (full checkpoint or lora) bleeding strongly into "huge"? I don't think so, not with reasonable training data.