IP-Adapter settings: weight, which model, and why your character becomes a copy

IP-Adapter encodes a reference image and injects it alongside the text prompt. Too little and it is ignored; too much and you get the reference back, pose and background included. Weight - 0.4-0.6: style and general likeness, the prompt still controls the scene. - 0.6-0.8: strong likeness, workable for characters. - 0.9-1.0: near-copy. The reference's pose, framing and background leak in. If your new scene keeps coming back as the old scene, lower the weight before anything else. Which IP-Adapter The project ships several variants, listed at https://github.com/tencent-ailab/IP-Adapter - The base and "plus" models for general image prompting. - "plus-face" variants that focus on the face crop. - FaceID variants that use a face-recognition embedding; stronger identity, but they need an extra face-analysis dependency. Use the variant that matches your base model (SD 1.5 or SDXL); they are not interchangeable. Crop the reference Use a clean, well-lit, front or three-quarter view, cropped to the face (for face variants) or the full figure (for costume). Busy backgrounds in the reference become busy backgrounds in the output. Combine with ControlNet IP-Adapter decides who; an openpose ControlNet decides the pose. That pairing is the most reliable non-training method for "same character, new pose". In the Studio Hathor Studio's character reference is IP-Adapter. Upload the clearest picture of the character you have; one good reference beats three mediocre ones.