InstantID
An open-source AI portrait generator that runs locally, preserves identity from one photo, and works with SDXL without personal fine-tuning.
InstantID generates images that retain a person's facial identity from a single reference photo, with text prompts controlling the style and scene. It's for creators and developers who want personalized avatars or portraits without collecting a photo set or training a model for each person. The Python code runs locally and includes a Gradio demo.
Its main distinction is the combination of face resemblance and prompt control. You can ask for realistic or stylized images, change poses, and blend the face into the generated scene. InstantID uses facial information and landmarks to guide a diffusion model while leaving the underlying model's image generation abilities intact.
The released local examples use SDXL through Hugging Face Diffusers. It also works with existing ControlNets, including canny and depth controls, for users who need more control over composition. Multiple reference photos can contribute to the identity representation, though a single image is sufficient. LCM-LoRA and OneDiff provide options for faster generation, and InstantStyle is compatible for style transfer.
The code is open source under Apache 2.0, which permits academic and commercial use. The InsightFace face models it relies on have a separate restriction: they're licensed for non-commercial research only.
