Nowadays almost everyone in the AI community talks about scaling laws. While many people believe AGI will go through scaling alone, there are also critics argue the opposite. In such cases, it’s better for us to understand scaling laws in details before joining the debate. In this talk, we will start by introducing the study on the scaling laws for language models, where the loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude. After that, we will move from language models to diffusion transformers (DiT) used for image generation, since DiTs are getting bigger and bigger and larger DiTs result in better visual quality and improved text-image alignment, yet the scaling law for DiT is less explored. Through experiments across a broad range of compute budgets, we confirm that the loss of pre-training DiT also follows a power-law relationship with the involved compute, which provides a predictable benchmark that assesses model performance and data quality at a reduced cost.