BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Semantic Information Encoded in Diffusion Models
DTSTART:20230717T093000
DTEND:20230717T113000
DTSTAMP:20260406T183405Z
UID:412128fac17a6b8f94d1f061c14f88ea4b5ce0b24a6980a685d5e5fa
CATEGORIES:Conferences - Seminars
DESCRIPTION:Shuangqi Li\nEDIC candidacy exam\nExam president: Prof. Alexan
 dre Alahi\nThesis advisor: Prof. Mathieu Salzmann\nCo-advisor: Prof. Sabin
 e Süsstrunk\nCo-examiner: Prof. Volkan Cevher\n\nAbstract\nDiffusion mode
 ls have achieved phenomenal success thanks to their ability to generate hi
 gh-quality images conditioned on text prompts. However\, to enable more co
 ntrol and flexibility in the generation process\, we are encouraged to del
 ve into the network architecture of diffusion models and investigate how s
 emantic information of generated images is stored and can be modified. In 
 this paper\, we begin by introducing DDPM (Denoising Diffusion Probabilist
 ic Models) proposed by~\\cite{ho2020denoising}\, which forms the backbone 
 of state-of-the-art diffusion models. We then present the idea of utilizin
 g the \\textit{h}-space as the semantic latent space in diffusion models\,
  as proposed by~\\cite{kwon2022asyrp}. Next\, we present the idea of Promp
 t-to-Prompt~\\cite{hertz2022prompt}\, which edits the interpretable cross-
 attention maps throughout the generation process in text-conditional diffu
 sion models. Furthermore\, we identify some challenges related to semantic
  control in diffusion models like Stable Diffusion. Finally\, we propose a
  method of editing self-attention maps and a method of guiding attention v
 ia large language models and showcase preliminary results.\n\nBackground p
 apers\nDenoising Diffusion Probabilistic Models\nDiffusion Models already 
 have a Semantic Latent Space\nPrompt-to-Prompt Image Editing with Cross-At
 tention Control.
LOCATION:BC 229 https://plan.epfl.ch/?room==BC%20229
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR