Diffusion models have become a cornerstone of generative AI, but choosing the right prediction target—epsilon, sample (x0), or v prediction—can significantly impact training stability and output quality. This post offers a straightforward comparison of these three approaches, explaining when each is most effective. Epsilon prediction is the classic choice for image generation, sample prediction simplifies certain loss calculations, and v prediction offers improved stability for high-resolution outputs. While the content is not novel, it serves as a useful reference for engineers implementing or fine-tuning diffusion models. The trade-offs between these targets are critical for production systems, making this a valuable evergreen resource for the AI community.
A clear breakdown of epsilon, sample, and v prediction in diffusion models, helping practitioners choose the right target.