7.3 Diffusion Models asoslari

📌 Ushbu darsda siz quyidagilarni o'rganasiz: 1️⃣ Forward Process — rasmga asta-sekin shovqin qo'shish va nima uchun 1000 qadam kerak 2️⃣ Reverse Process — shovqindan rasmni qayta tiklash va U-Net rolini tushunish 3️⃣ Score Matching — log ehtimollik gradienti va shovqin bashorati bog'liqligi 4️⃣ DDPM Loss — murakkab VLB qanday qilib oddiy MSE ga aylandi 5️⃣ U-Net + Time Embedding — skip connections va sinusoidal embedding nima uchun muhim 🎯 Bu dars orqali siz quyidagini chuqur tushunasiz: GAN rasmni "ixtiro qiladi", VAE siqib qayta tiklaydi. Diffusion esa shovqindan asta-sekin rasm "quradi" — va bu eng sifatli natija beradi. Bugungi eng kuchli generativ modellari — Stable Diffusion, DALL-E 2, Midjourney, Imagen — barchasi DDPM asosida qurilgan. 👉 Kurs rejasi (to'liq): 🔗   / deep-learning-matematikasi-intensiv-kurs-r...   📌 Telegram kanal: 👉 https://t.me/EldorML