Keep a character consistent

tendre.red keeps a character's face and identity stable across images. How it does that depends on how the character was created.

Reference image, IPAdapter and InstantID

When you create a character with Prompt to Image (from traits) or Image to image (from 1 to 3 photos), tendre.red stores a reference image, a base checkpoint and, for trait-based characters, a fixed seed on that character's record. Every new generation reuses the reference image through IPAdapter, and InstantID (an SDXL face-consistency model) reinforces the face on top of it. No training needed, and it works from a single photo.

When to train a LoRA instead

For maximum fidelity, especially full-body shots or heavier changes in pose, outfit and lighting, train a LoRA on 10 to 15 photos (see Train a LoRA). Once trained, the LoRA generation tab carries the identity through the LoRA itself, using the character's trigger word, without InstantID in that mode.

Prompt around the character

Whichever method you used, describe the scene, outfit and pose as usual. The face stays the same person while everything else in the prompt changes freely.

Tips for best results

Keep the checkpoint you trained or created the character on, switching base models can shift the likeness. If a character keeps drifting, revisit your photo dataset and consider training or retraining its LoRA.