Diffusers for Democratising Diffusion Models

Sayak Paul, Developer Advocate Engineer, Hugging Face

12 June 2023

Talk summary: In diffusion models, noise vectors are continuously refined to create realistic images. However, diffusers are used for applications beyond image generation. At Hugging Face, diffusers are used in image translation, text to video generation, latent space manipulation, image editing with human-readable instructions, and semantic guidance. These diffusers have practical applications in the comics industry and in marketing, to name a few.

Not all diffusion models are open source. Why do we need to make them open source? This is important to study risk factors and failure cases, evaluate safety measures, build, and improve on them. The Hugging Face team makes diffusers as accessible as possible to researchers. The Python library maintained at https://lnkd.in/dCj9ubbv provides open and responsible access to pre-trained diffusion models. The aim is to democratise the ecosystem of diffusion models by making them easy to use. Sayak Paul emphasised that the Hugging Face team is open to collaborations.

Speaker bio: Sayak Paul works on the Diffusers team at Hugging Face, where he codes, trains, and documents diffusion models.

[Talk organised in collaboration with the Department of Computational and Data Sciences]