FaceChain-FACT

FaceChain-FACT is a zero-shot portrait generation model developed by Alibaba Group that does not require Face LoRA model training. It uses a single photo of a user to generate customized portraits with second-level inference speed, accelerated by 100 times compared to SOTA commercial applications. The model integrates a transformer-based face feature extractor with a structure similar to Stable Diffusion, enabling better utilization of face information through dense fine-grained features as face conditions. FaceChain-FACT is seamlessly compatible with ControlNet and LoRA plugins, allowing plug-and-play usage. It supports CPU inference and offers over 100 haute couture templates.