<ul data-eligibleForWebStory="true"><li>Researchers have introduced Ambient Diffusion Omni, a framework to train diffusion models using low-quality, synthetic, and out-of-distribution images.</li><li>Traditional diffusion models are trained on curated datasets but the new approach aims to leverage lower-quality images usually discarded.</li><li>The framework leverages spectral power law decay and locality properties of natural images.</li><li>They successfully trained diffusion models with synthetically corrupted images and achieved state-of-the-art ImageNet FID.</li><li>Significant improvements in image quality and diversity were observed for text-to-image generative modeling.</li><li>The core insight of the framework is that noise helps in learning from biased data and mixed distributions.</li><li>The approach was validated through the trade-off analysis between biased but abundant data and limited unbiased data during diffusion times.</li></ul>

Ambient Diffusion Omni: Training Good Models with Bad Data

Discover more