Скачать или смотреть AI Image Generation Just Got a HUGE Speed Boost! SD3 & FLUX Distilled

We all love the incredible art that AI can generate, but the slow speeds have always been a major hurdle. Top-tier models like Stable Diffusion 3 often require dozens of steps to create a single image, which takes time and serious computing power.

Historically, AI researchers have used two main approaches for image generation: 'diffusion' and 'flow matching'. These were often treated as separate worlds, with different tricks needed to speed them up. This new research paper bridges that gap, showing that these two methods are fundamentally two sides of the same coin.

Building on this insight, the researchers applied a powerful technique called Score Distillation to the latest flow matching models, shrinking them down into hyper-efficient versions that can generate stunning images in just four steps! This video breaks down how they did it, what it means for models like SD3 and FLUX, and why this could be a game-changer for the future of accessible, real-time AI generation.

Cited paper:
M. Zhou et al. (2025). Score Distillation of Flow Matching Models. arXiv:2509.25127v1. http://arxiv.org/abs/2509.25127v1

Images shown are page renders from the paper PDF for commentary/education.