Following SDXL-Lightning, ByteDance reaffirms its commitment to innovation and speed with the launch of Hyper-SD, a new technique for optimizing and accelerating the speed of image generation with diffusion models.
ByteDance’s goal seems to be to move even closer to real-time image generation while maintaining the best possible quality. With this in mind, Hyper-SD represents a new step forward and generates better quality images than SDXL Turbo or SDXL-Lightning, further improving generation speed.
Like its predecessors, Hyper-SD is available in the form of LoRAs compatible with Stable Diffusion models, which can be used with between 1 and 8 inference steps. The files are freely downloadable from Hugging Face.
Hyper-SD uses a dual approach: two-step Trajectory Segmented Consistency Distillation, which maintains the integrity of the original ODE (Ordinary Differential Equation) trajectory, and Learning with Human Feedback, which improves image quality.