Character LoRA bleed-through in ComfyUI: why one face overwrites the whole scene

A ComfyUI user posted a workflow screenshot this week showing a character LoRA bleeding into every face in a multi-subject scene, even when the prompt explicitly called for different characters. The LoRA—trained on a single anime character—overwrote all facial features in the output, turning what should have been a group shot into a repetition of the same face. The issue is familiar to anyone who's tried to mix character LoRAs with generic prompts: the model latches onto the trained identity and ignores compositional cues.

The most common workaround is dialing the LoRA strength below 0.5, but that often dilutes the character features to the point of unrecognizability. Some workflows layer regional conditioning nodes to mask the LoRA's influence to specific bounding boxes, effectively telling the sampler "apply this face here, not everywhere." Others inject negative embeddings of the character name into non-target regions, though that requires trial-and-error tuning per seed. A third approach chains two separate samplers—one with the LoRA active for the main subject, a second pass with it disabled for background figures—but that doubles generation time and can introduce seam artifacts where the masks meet.

The bleed-through stems from how LoRAs modify the base model's attention layers: a character-specific LoRA trained on 50–100 images of one face will push the entire latent space toward that identity, especially in lower-resolution or crowded compositions where the model has less spatial information to differentiate subjects. Fine-tuning the LoRA with more varied backgrounds and secondary characters during training can reduce the effect, but most publicly shared character LoRAs are optimized for solo portraits, not ensemble scenes. Until ComfyUI ships native per-region LoRA routing or base models gain better compositional grounding, multi-character workflows will remain a game of strength sliders and mask geometry.