This workflow uses two custom nodes to recreate a character’s identity without relying on an edit LoRA or other LoRAs. Don’t expect a perfect match, but it can preserve the character surprisingly well across different scenes, outfits, and styles.

I recommend starting with medium identity detail, as it usually gives the best balance between identity preservation and prompt flexibility. Higher settings strengthen the identity reference but can reduce scene and framing freedom.

For automatic face detection/cropping, you’ll need MediaPipe and opencv-python-headless. The workflow can still be used without them, but the automatic crop modes will not work.

The zip includes the nodes and the workflow.

Bonus: Depending on the settings, you can also use the workflow to recreate or reinterpret input images. See the examples in the gallery.