Switti Switti is a groundbreaking research project that introduces a scale-wise transformer for text-to-image synthesis. The team behind Switti, consisting of Anton Voronov, Denis Kuznedelev, Mikhail Khoroshikh, Valentin Khrulkov, and Dmitry Baranchuk, has taken existing next-scale prediction AR models and enhanced them for T2I generation. By making architectural modifications to improve convergence and overall performance, Switti demonstrates significant advancements in this field.
One of the key findings in their research is the weak dependence of self-attention maps in their pretrained scale-wise AR model on preceding scales. Leveraging this insight, they propose a non-AR counterpart that not only accelerates sampling by approximately 11% but also reduces memory usage while maintaining generation quality. Additionally, they discovered that guidance at high-resolution scales may be unnecessary and can even hinder performance. By disabling guidance at these scales, Switti achieves a further 20% acceleration in sampling and enhances the generation of fine-grained details.
Human preference studies and automated evaluations have shown that Switti surpasses existing T2I AR models and competes with state-of-the-art diffusion models while being up to 7 times faster. The team’s meticulous evaluation process includes human evaluations, inference performance evaluations, and automated metrics, all of which highlight the superior performance of Switti in comparison to its counterparts.
Their research, titled “Switti: Designing Scale-Wise Transformers for Text-to-Image Synthesis,” presents a comprehensive analysis of their methodology and findings. The team’s work is not only innovative but also highly impactful in the realm of text-to-image synthesis, pushing the boundaries of what is achievable in this field.
Visit Site