You can achieve precise character preservation without training LoRA

A ComfyUI node package called Flux2Klein-Enhancer has done exactly one thing—when generating images with the Flux model, it keeps a character’s appearance precisely and consistently without any additional LoRA training.
In the past, for AI face swapping, AI portraits, and IP character consistency, you had to first train a LoRA or use an IP-Adapter. Training time, compute resources, and fine-tuning—the bar was not low.
This approach goes straight around all of that.
The principle is: in the inference stage, do feature injection and enhancement, and by leveraging the Flux model’s own capabilities, it achieves fidelity close to LoRA-level performance.
After looking at the updated results comparison, the consistency of facial details, hairstyles, and skin tones across different scenes really is noticeably better than the previous version.
The toolchain is getting shorter, and the creative barrier is getting lower. The remaining competition is only in aesthetics and storytelling.