SD3.5-Flash: Distribution-Guided Distillation for Generative Flow

SD3.5-Flash is an efficient distillation framework that compresses computationally expensive rectified flow models to run in just a few steps. The key is reformulating a distribution matching objective tailored for few-step generation, enabling fast execution on consumer devices while preserving quality.

There are two main improvements.

• timestep sharing: reduces gradient noise to stabilize training
• split-timestep fine-tuning: improves prompt alignment

On top of this, pipeline optimizations such as text encoder restructuring and specialized quantization enable memory-efficient deployment across diverse hardware, from mobile phones to desktops. The result points toward a direction that delivers advanced generative AI to a broader user base.

In evaluations including a large-scale user study, SD3.5-Flash is presented as consistently outperforming existing few-step methods. In other words, this isn't just a fast model—the focus is on achieving both the speed and quality needed for real-world use together.