Translated data: According to Zhanzhang Zhijia, InstaFlow is a model that generates images in a single step, leveraging Rectified Flow technology to train a probabilistic flow along a straight trajectory, thus requiring only one inference step. Compared to Stable Diffusion, it can reduce inference time by 90% while maintaining similar image quality. InstaFlow employs supervised training, generating text-noise-image triplets through a pre-trained Stable Diffusion model, and then applying text-adjusted backflow to produce a straight probabilistic flow. Finally, it transfers learning from this to obtain the one-step generator InstaFlow. This model can quickly generate high-quality images, representing a new generation of text-to-image generation models.