InstantID

InstantID, from the InstantX team at Xiaohongshu (paper arXiv 2401.07519, January 2024), brought zero-shot identity preservation to the SDXL era. Feed it one reference portrait and a text prompt, and it generates a new image of that person in a different scene, style, or outfit, with no per-identity training. It works by combining a face ID embedding (from an InsightFace model) with five-point facial landmarks fed through an IdentityNet ControlNet, which anchors the pose and structure while the ID adapter carries the likeness. Generation takes seconds on a single GPU.

The code is Apache 2.0, but the released checkpoints are research-only and the InsightFace face models are non-commercial, a licensing split that matters for anyone shipping a product on top of it. InstantID exploded in the community: ComfyUI nodes, WebUI/ControlNet integration, Replicate demos, and one-click Windows builds, and it underpins a large share of the personalized-avatar workflows on Civitai. There are no published standard benchmarks; the paper compares qualitatively against IP-Adapter and ControlNet baselines. Later work like InfiniteYou and PuLID built on or competed with the same idea, but InstantID remains the most deployed implementation of the technique.

No published standard benchmark scores; paper compares qualitatively against IP-Adapter and ControlNet baselines