STABLE DIFFUSION
SD Turbo — a fast generative text-to-image model base on SD 2.1
SD Turbo is a distilled version of Stable Diffusion 2.1, trained for real-time synthesis.
The foundation of SD Turbo lies in a groundbreaking training method known as Adversarial Diffusion Distillation (ADD) (refer to the technical report). This innovative approach enables the extraction of extensive foundational image diffusion models within 1 to 4 steps while maintaining high image quality. By utilizing score distillation, it harnesses pre-existing large-scale image diffusion models as a guiding signal and integrates an adversarial loss component to ensure exceptional image fidelity, even when employing only one or two sampling steps.
The charts provided offer insights into user preferences regarding SD-Turbo compared to both single- and multi-step models. Specifically, when evaluated at a single step, SD Turbo is favored by human voters in terms of image quality and prompt adherence over alternatives like LCM-Lora XL and LCM-Lora 1.5.
While SD Turbo contributes to improved generation speed, its foundation lies in SD 2.1. However, SD 2.1 is considerably less popular than both SD 1.5 and SDXL, resulting in SD Turbo receiving comparatively less attention. For those seeking higher image quality, opting for SDXL Turbo is recommended.