eDiffi: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers

Proposes an ensemble of text-to-image diffusion models specialized for different synthesis stages to improve text alignment while maintaining high visual quality and the same inference computation cost. Shows improved results compared to previous large-scale text-to-image diffusion models on standard benchmarks. Enables 'paint-with-words' capability to allow users to select words in input text and paint them in a canvas to control the output.

Can improve text alignment and visual quality of text-to-image synthesis in business applications, such as generating product images or visuals for marketing campaigns. 'Paint-with-words' capability can be useful for designers and marketing professionals to create desired visuals.