Latent diffusion models (LDMs) are studied for their scaling properties, particularly in terms of sampling efficiency. The paper investigates how model size affects sampling efficiency across different steps. It finds that smaller models often outperform larger ones under the same inference budget, challenging the assumption that larger models are always better. The study also explores the generalizability of these findings across various tasks and diffusion samplers, showing that smaller models maintain superior sampling efficiency regardless of the sampler used. Additionally, the research demonstrates that smaller models can achieve better results with fewer sampling steps, and that diffusion distillation does not fundamentally alter the scaling trends. The findings suggest that LDMs can be scaled efficiently to enhance generative capabilities within limited inference budgets. The study trains a range of LDMs from 39 million to 5 billion parameters, showing that smaller models can achieve better performance in terms of FID scores and visual quality, especially under constrained sampling costs. The results indicate that scaling LDMs can lead to improved performance without necessarily increasing model size, offering new insights into the development of efficient LDM scaling strategies.Latent diffusion models (LDMs) are studied for their scaling properties, particularly in terms of sampling efficiency. The paper investigates how model size affects sampling efficiency across different steps. It finds that smaller models often outperform larger ones under the same inference budget, challenging the assumption that larger models are always better. The study also explores the generalizability of these findings across various tasks and diffusion samplers, showing that smaller models maintain superior sampling efficiency regardless of the sampler used. Additionally, the research demonstrates that smaller models can achieve better results with fewer sampling steps, and that diffusion distillation does not fundamentally alter the scaling trends. The findings suggest that LDMs can be scaled efficiently to enhance generative capabilities within limited inference budgets. The study trains a range of LDMs from 39 million to 5 billion parameters, showing that smaller models can achieve better performance in terms of FID scores and visual quality, especially under constrained sampling costs. The results indicate that scaling LDMs can lead to improved performance without necessarily increasing model size, offering new insights into the development of efficient LDM scaling strategies.