← Back to Multimodal Mind Map
👁️ MultimodalID: diffusion-vs-gan

Diffusion vs GAN vs Autoregressive

扩散 vs GAN vs 自回归
🎯Core Definition
The three major generation families trade off quality/diversity/speed. GANs play a zero-sum game between generator and discriminator — one forward pass is fastest, but adversarial training is unstable and prone to mode collapse. Diffusion/flow models generate by progressive denoising: training is stable regression and diversity is strong, and current SOTA FID is lower (higher quality ceiling). Autoregressive models predict token by token — great for text and discrete modalities with natural conditioning, but sampling is serial. Comparison:
📌Overview
| Dimension | GAN | Diffusion (DDPM/distilled) | Autoregressive | |---|---|---|---| | Training stability | poor (adversarial) | good (MSE regression) | good (cross-entropy) | | Diversity | poor (mode collapse) | good (full coverage) | good | | Quality (FID) | realistic in-mode | SOTA lowest FID | weaker on images | | Sampling speed | 1 step (fastest) | 4–1000 steps (DDIM/distillable) | serial tokens | | High-res scaling | unstable | latent + DiT scalable | hard |
📌Overview
FID (Fréchet Inception Distance): pass real and generated images through an Inception network, assume the features are Gaussian, and measure the Fréchet distance between their means and covariances:
📌Overview
FID=∥μr−μg∥2+Tr(Σr+Σg−2(ΣrΣg)1/2)FID = \Vert \mu_r - \mu_g \Vert^2 + \mathrm{Tr}\left(\Sigma_r + \Sigma_g - 2(\Sigma_r\Sigma_g)^{1/2}\right)
📌Overview
Lower is better; it captures both fidelity and diversity, though note they are not the same — a GAN can have decent FID yet poor diversity.
💡Use Cases
interview comparison questions (“why is text-to-image SOTA diffusion, not GAN?”); architecture selection — low-latency interactive tasks (real-time filters, SR) favor GANs or distilled diffusion, quality/diversity-first tasks (text-to-image, video) favor diffusion.
⚡Key Problems Solved
GANs' unstable adversarial training, mode collapse and hard high-resolution scaling; diffusion trades more sampling steps for stable training, full distribution coverage and lower FID; autoregressive models have been largely displaced for image generation by diffusion/flow models.
🎯Core Interview Exam Points
Fundamentals · Production Scenarios · Deep Dive
Core Def
Explain the core definition, underlying mechanism/math, and key principles of Diffusion vs GAN vs Autoregressive.
Use Cases
In production engineering, what are the primary use cases for Diffusion vs GAN vs Autoregressive and what core pain points does it solve?
🔥 High-Frequency Deep-Dive Questions
1
Compare diffusion vs GAN on quality, diversity and speed; why is diffusion the T2I SOTA?
2
Write and explain the FID formula; why does it not fully capture diversity?
3
Causes and detection of GAN mode collapse; how does diffusion naturally avoid it?
4
Pros and cons of autoregressive vs diffusion generation; why is AR being displaced on images?
5
For real-time use (mobile SR/filters), GAN or distilled diffusion — how to trade off latency and quality?
📖 In-depth Guide:📄 diffusion-models →
Updated 2026-08-12
🎯
Verify Mastery: Practice Exam Questions for "Diffusion vs GAN vs Autoregressive"
Take pitfall multiple-choice questions with derivation options; mistakes are automatically logged.
🚀 Start Card Practice ➔
← Previous CardDiT & 3D Causal VAENext Card →Diffusion Training Objectives (ε/v-pred, Zero-SNR)

🔗 More Multimodal Knowledge Cards

引用出处: TalentMe 450+ 工业级 AI 架构考点知识库 (Tech Vault) · 遵循 CC-BY-4.0 知识共享协议
URL: https://talentme.airsota.com/resources/tech/roadmap/mindmap/Multimodal/card/diffusion-vs-gan
中文速览与面试要点

扩散 vs GAN vs 自回归 (Diffusion vs GAN vs Autoregressive)

定义: 生成模型三大路线的对比, 核心是质量/多样性/速度三角。GAN 用生成器与判别器的零和博弈, 单步前向推理最快, 但对抗训练不稳定、易模式坍塌; 扩散/流模型通过逐步去噪生成, 训练是稳定回归、多样性好, 当前 SOTA FID 更低(质量上限更高); 自回归逐 token 预测, 擅长文本与离散模态、天然支持条件化, 但生成必须串行。对比表: | 维度 | GAN | 扩散 (DDPM/蒸馏) | 自回归 | |---|---|---|---| | 训练稳定性 | 差 (对抗不收敛) | 好 (MSE 回归) | 好 (交叉熵) | | 多样性 | 差 (模式坍塌) | 好 (覆盖全分布) | 好 | | 质量 (FID) | 模式内逼真 | 当前 SOTA 最低 FID | 图像上较弱 | | 采样速度 | 1 步 (最快) | 4–1000 步 (可蒸馏/DDIM) | 逐 token 串行 | | 高分辨率扩展 | 难稳定 | 潜空间 + DiT 可扩展 | 难 | FID (Fréchet Inception Distance): 把真实与生成图像分别过 Inception 网络取特征, 假设特征服从高斯分布, 计算两组均值与协方差之间的 Fréchet 距离: FID = _r - _g ^2 + Tr ( _r + _g - 2( _r _g)^1/2 ) 越低越好, 同时衡量真实度与多样性; 注意 FID 与多样性并不完全等价——GAN 可能 FID 尚可但多样性差。 使用场景: 面试对比题 (为什么文生图 SOTA 是扩散而非 GAN?); 生成方案选型: 低延迟交互 (实时滤镜、SR 超分) 用 GAN 或蒸馏扩散, 质量/多样性优先 (文生图、视频) 用扩散。 解决的问题: GAN 的对抗训练不稳定、模式坍塌与高分辨率难扩展; 扩散以更多采样步数为代价换来稳定训练、完整分布覆盖与更低 FID; 自回归在图像生成上的位置逐渐被扩散/流模型取代。

核心要点

  • 扩散 vs GAN 在质量、多样性、速度上的对比? 为什么文生图 SOTA 用扩散?
  • 写出 FID 公式并解释其含义, 为什么它不能完全代表多样性?
  • GAN 模式坍塌的成因与检测方法? 扩散如何天然规避这一问题?
  • 自回归生成相比扩散的优缺点? 为什么自回归在图像上逐渐被取代?
  • 实时场景(如手机端超分/滤镜)选 GAN 还是蒸馏扩散? 各自延迟与质量如何权衡?