Switti

Switti is a scale-wise transformer architecture for text-to-image synthesis, designed by Yandex Research and presented at CVPR 2025. It adapts next-scale prediction autoregressive (AR) architectures to T2I generation while addressing training stability issues. The model introduces a non-causal transformer variant that enables approximately 21% faster sampling and lower memory usage with slightly better generation quality. Additionally, it reveals that classifier-free guidance at high-resolution scales is often unnecessary and can degrade performance; disabling guidance at these scales yields an additional ~32% acceleration and improves fine-grained detail generation. Switti achieves state-of-the-art results among AR models and competes with diffusion models while being up to 7× faster in generating 1024×1024 images.