IP-Adapter

IP-Adapter from Tencent AI Lab lets pretrained text-to-image diffusion models accept images as prompts, alone or combined with text. The adapter adds only about 22 million parameters, an image projection module plus decoupled cross attention layers, so a reference image can steer generation with quality comparable to a fully fine-tuned image prompt model while the base model stays frozen. It works with Stable Diffusion 1.5, SDXL, and community checkpoints derived from them, uses an OpenCLIP ViT-H-14 image encoder, and composes with ControlNet and T2I-Adapter for structural control. Specialized FaceID variants target identity preserving portrait generation, and image guided inpainting is supported. Integrations in Hugging Face Diffusers, the AUTOMATIC1111 WebUI, ComfyUI, and InvokeAI have made it a standard building block for style transfer, image variation, and subject driven workflows. The code is open source under the Apache 2.0 license.