InstantID

InstantID generates identity-preserving images from a single reference face photo with no per-subject training or fine-tuning. The tuning-free pipeline extracts facial embeddings and keypoints with InsightFace's AntelopeV2 model, conditions generation through IdentityNet, a ControlNet-style network driven by facial landmarks, and injects the identity embedding via an IP-Adapter module on top of a Stable Diffusion XL base. The result places a specific person into new styles and scenes from a text prompt while keeping the face recognizable, and it works with community SDXL checkpoints as well as LCM-LoRA for accelerated generation in around 10 steps. Developed by the InstantX team, the code is Apache 2.0 licensed, though the bundled InsightFace models are restricted to non-commercial research use. Adoption spans Hugging Face Spaces demos plus ComfyUI and WebUI integrations, and it is widely used for stylized portraits, avatars, and character-consistent image generation.