Contents

Denoising Diffusion Probabilistic Models (DDPM)

arXiv 2006.11239 Hugging Face

TL;DR

Motivations & Innovations

Approach

Model

normal gaussian distribution:

./images/index-20260314201005.webp

./images/index-20260314201327.webp

./images/index-20260314201504.webp

./images/index-20260314201628.webp

加噪过程 -> 生产数据

./images/index-20260314201726.webp

加入 text control

./images/index-20260314202113.png

训练

./images/index-20260314202210.webp

data

./images/index-20260314201958.webp

laion data browser

算法细节

./images/index-20260314202238.webp

forward process:

reverse process

./images/index-20260314203004.webp

./images/index-20260314203257.webp

VAE vs. Diffusion Model

./images/index-20260314202407.webp

Diffusion-based Products

Diffusion 模型在图像生成领域取得了巨大成功,催生了许多商业化产品和知名公司:

产品/公司特点链接
DALL·E (OpenAI)集成于 ChatGPT,支持 API 调用,强调对文本提示的精准遵循OpenAI
Stable Diffusion (Stability AI)开源模型,2024年发布 3.5 版本 (Large/Large Turbo/Medium),社区版免费商用Stability AI
Midjourney (独立研究实验室)以艺术风格见长,通过 Discord 社区运营Midjourney
Imagen (Google DeepMind)2026年推出 Nano Banana 2 (Gemini 3.1 Flash Image),支持实时生成Google DeepMind
Adobe Firefly (Adobe)集成于 Creative Cloud,主打商业安全合规Adobe Firefly