DDPMs - Denoising Diffusion Probabilistic Models
In this video, I theoretically explained how DDPM models work. This is NOT A CODING video.
In simple terms, two processes happen in denoising diffusion probabilistic models (𝐃𝐃𝐏𝐌𝐬). First, The data structure is destroyed by gradually adding Gaussian noise over a finite number of time steps to end up with pure noise (𝐟𝐨𝐫𝐰𝐚𝐫𝐝/𝐝𝐢𝐟𝐟𝐮𝐬𝐢𝐨𝐧 𝐩𝐫𝐨𝐜𝐞𝐬𝐬). Second, A neural network is trained to gradually denoise the data starting from pure noise and predict a distribution that looks like the original distribution (𝐫𝐞𝐯𝐞𝐫𝐬𝐞/𝐝𝐞𝐧𝐨𝐢𝐬𝐢𝐧𝐠 𝐩𝐫𝐨𝐜𝐞𝐬𝐬).
𝗗𝗗𝗣𝗠 𝗺𝗼𝗱𝗲𝗹𝘀 𝗮𝗿𝗲 𝗲𝘅𝗽𝗲𝗰𝘁𝗲𝗱 𝘁𝗼 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝗮𝗰𝘁𝘂𝗮𝗹 𝗶𝗺𝗮𝗴𝗲𝘀 𝗳𝗿𝗼𝗺 𝗻𝗼𝗶𝘀𝗲 𝗼𝗻𝗹𝘆 𝗶𝗳 𝘁𝗿𝗮𝗶𝗻𝗲𝗱 𝘄𝗲𝗹𝗹.
𝙂𝙞𝙩𝙃𝙪𝙗 𝙖𝙙𝙙𝙧𝙚𝙨𝙨: https://github.com/randomaccess2023/M...
#generativemodels #generativeai #diffusionmodels #DDPM #deeplearning