Ambient Diffusion Omni: Training Good Models with Bad Data

Summary

This paper talks about Ambient Diffusion Omni, a new framework that trains image generation models using all kinds of images, including low-quality, blurry, or synthetic ones, instead of only clean, high-quality images.

What's the problem?

The problem is that most image generation models are trained only on carefully selected, clean images, which limits the amount of data available and can make models less diverse and less adaptable to real-world images that are not perfect.

What's the solution?

The researchers designed Ambient Diffusion Omni to use properties of natural images to figure out how to best use every image during training, even if it is noisy or corrupted. This approach lets the model learn from a bigger and more varied set of images by adjusting how it processes each one based on its quality and noise level.

Why it matters?

This matters because using all available images, including lower-quality ones, helps AI models generate better and more diverse images. It also reduces the cost and effort of collecting perfect images and makes the models more robust to different kinds of inputs.