PhotoMaker

PhotoMaker, from Tencent ARC Lab and Nankai University, generates customized realistic photos of a specific person from a handful of reference images, with no per-subject LoRA training required. Its core technique is a stacked ID embedding that fuses identity information from several input photos, letting the model preserve who the person is while text prompts change style, scene, age, or composition, all within seconds on a single GPU with roughly 11 GB of memory. Built on Stable Diffusion XL, it functions as an adapter compatible with other base models, LoRA modules, ControlNet, T2I-Adapter, and IP-Adapter. The work was published at CVPR 2024, and a V2 release in July 2024 improved identity fidelity while retaining editability and generation quality. Code and weights are open under the Apache 2.0 license, with integrations for ComfyUI, the AUTOMATIC1111 WebUI, Replicate, and Colab, and it is widely used for portrait generation and stylized avatars.