Internet video streaming has experienced explosive growth over the past few years. Recently, super-resolution (SR) networks have been utilized to reduce the bandwidth and improve the quality of Internet video streaming. These methods first employ a predefined and immutable downscaling kernel, such as bicubic interpolation, to transform high-resolution (HR) video into low-resolution (LR) video. Subsequently, the LR video is partitioned into segments, which are streamed alongside corresponding models to the clients. The client subsequently executes inference models to perform SR on the LR segments. However, this normal downscaling is not an injective mapping because high-frequency information is lost. This creates the ill-posed problem of the inverse upscaling procedure and makes it highly difficult to get details back from downscaled LR videos. In this paper, we propose a novel method for video delivery. Specifically, we deliberately designed an Enhanced Asymmetric Invertible Network (EAIN) to produce high-quality LR videos while capturing the distribution of missing information using a latent variable that follows a specified distribution in the downscaling process. HR videos are available by passing a randomly extracted latent variable through the network in reverse with LR videos. Extensive experiments show that our methods significantly improve video streaming quality compared to state-of-the-art neural video delivery methods, paving the way for the application of neural video delivery techniques in practice. The code is available at https://github.com/Anonymous-ACCV-2024/EAIN .