第21章：CFG — Classifier-Free Guidance | Chapter 21: CFG — Classifier-Free Guidance

阶段定位 | Stage: 第四阶段 — Diffusion 预计学时 | Duration: 3~5 小时

---

学习目标 | Learning Objectives

中文：

• 理解 CFG 的核心思想：用"有条件"减去"无条件"
• 掌握 CFG 的公式和实现
• 理解为什么 CFG 越高越"听 prompt"
• 了解 CFG 的副作用和最佳实践

English:

• Understand core idea of CFG: "conditional" minus "unconditional"
• Master CFG formula and implementation
• Understand why higher CFG means more "prompt-following"
• Learn about CFG side effects and best practices

---

21.1 为什么需要 CFG？| Why CFG?

中文解释

问题：模型生成的图片不够"听"prompt

原因：

• 训练时模型学了各种图片，不只是 prompt 对应的图片
• 生成时会走"平均"路线，导致图片模糊、不精确

CFG 解决：放大条件的影响

English Explanation

Problem: Generated images don't "follow" prompt closely enough

Reason:

• During training, model learns all kinds of images, not just prompt-matching ones
• During generation, it takes an "average" route, causing blurry, imprecise images

CFG solution: Amplify the effect of conditioning

---

21.2 CFG 公式 | CFG Formula

中文解释

原始预测：| Original prediction:
  ε_cond — 有条件（有 prompt）的噪声预测 | with prompt

无条件预测：| Unconditional prediction:
  ε_uncond — 无条件（空 prompt）的噪声预测 | with empty prompt

CFG 后的预测：| CFG-adjusted prediction:
  ε_cfg = ε_uncond + scale × (ε_cond - ε_uncond)
        = (1 - scale) × ε_uncond + scale × ε_cond

当 scale = 1：就是有条件预测 | When scale = 1: just conditional prediction 当 scale > 1：放大条件的影响 | When scale > 1: amplify conditioning effect

English Explanation

Original prediction:
  ε_cond — noise prediction with prompt

Unconditional prediction:
  ε_uncond — noise prediction with empty prompt

CFG-adjusted prediction:
  ε_cfg = ε_uncond + scale × (ε_cond - ε_uncond)
        = (1 - scale) × ε_uncond + scale × ε_cond

When scale = 1: just conditional prediction When scale > 1: amplify conditioning effect

---

21.3 CFG 实现 | CFG Implementation

代码案例 | Code Example

import torch

def apply_cfg(noise_cond, noise_uncond, scale=7.5):
    """
    应用 Classifier-Free Guidance | Apply Classifier-Free Guidance
    
    noise_cond: 有条件预测的噪声 | Conditional noise prediction
    noise_uncond: 无条件预测的噪声 | Unconditional noise prediction
    scale: guidance scale
    """
    # CFG 公式 | CFG formula
    noise_cfg = noise_uncond + scale * (noise_cond - noise_uncond)
    return noise_cfg

# 在采样循环中使用 | Use in sampling loop
@torch.no_grad()
def sample_with_cfg(model, prompt_embeds, uncond_embeds, shape, scale=7.5):
    """
    使用 CFG 的采样 | Sampling with CFG
    """
    x = torch.randn(shape, device=prompt_embeds.device)
    
    for t in reversed(range(num_timesteps)):
        t_batch = torch.full((shape[0],), t, device=x.device)
        
        # 有条件预测 | Conditional prediction
        noise_cond = model(x, t_batch, prompt_embeds)
        
        # 无条件预测 | Unconditional prediction
        noise_uncond = model(x, t_batch, uncond_embeds)
        
        # CFG
        noise_pred = apply_cfg(noise_cond, noise_uncond, scale)
        
        # 去噪一步 | Denoise one step
        x = denoise_step(x, noise_pred, t)
    
    return x

# 直观理解 | Intuitive understanding:
# scale = 0:  完全无条件生成 | Fully unconditional
# scale = 1:  正常条件生成 | Normal conditional
# scale = 7.5: 强烈遵循 prompt | Strongly follow prompt (SD default)
# scale = 15: 极强遵循，但可能过饱和 | Very strong, but may oversaturate

---

21.4 为什么 CFG 越高越"听 prompt"？| Why Higher CFG Means More "Prompt-Following"?

中文解释

ε_cfg = ε_uncond + scale × (ε_cond - ε_uncond)

(ε_cond - ε_uncond) = 条件带来的"方向"
                      = 从无条件指向有条件的方向

scale 越大，这个方向的步长越大：
  - 模型会更坚定地朝 prompt 描述的方向走
  - 但也会更远离"平均"图片
  - 过高会导致过饱和、不自然

English Explanation

ε_cfg = ε_uncond + scale × (ε_cond - ε_uncond)

(ε_cond - ε_uncond) = "direction" brought by conditioning
                      = direction from unconditional to conditional

Larger scale means larger step in this direction:
  - Model more firmly walks toward prompt-described direction
  - But also farther from "average" images
  - Too high causes oversaturation, unnatural results

可视化 | Visualization

噪声预测空间（简化 2D）| Noise prediction space (simplified 2D):

         ε_uncond（无条件 | unconditional）
           ↓
           ●
           │
           │ scale=1
           │
           ● ← ε_cond（有条件 | conditional）
           │
           │ scale=7.5
           │
           ★ ← ε_cfg（CFG 后 | after CFG）
           
★ 离 ε_uncond 更远，离 prompt 方向更近
★ is farther from ε_uncond, closer to prompt direction

---

21.5 CFG 的副作用 | Side Effects of CFG

中文解释

English Explanation

训练时的 CFG | CFG During Training

# 训练时需要同时学习有条件和无条件 | Training needs to learn both conditional and unconditional

def train_step_with_cfg(model, x_0, prompt_embeds):
    """
    训练时随机丢弃条件 | Randomly drop conditioning during training
    """
    batch_size = x_0.size(0)
    
    # 随机丢弃条件（概率 10%）| Randomly drop conditioning (10% probability)
    mask = torch.rand(batch_size) > 0.1
    
    # 被丢弃的用空 embedding | Use empty embedding for dropped ones
    context = prompt_embeds.clone()
    context[~mask] = 0   # 空条件 | Empty conditioning
    
    # 正常训练 | Normal training
    t = torch.randint(0, num_timesteps, (batch_size,))
    x_t, noise = diffusion.add_noise(x_0, t)
    pred_noise = model(x_t, t, context)
    
    loss = torch.nn.functional.mse_loss(pred_noise, noise)
    return loss

---

本章总结 | Chapter Summary

中文：

• CFG = ε_uncond + scale × (ε_cond - ε_uncond)
• scale 控制 prompt 的影响强度
• 训练时随机丢弃条件（10%）来学习无条件生成
• scale 默认 7.5，过高会导致过饱和

English:

• CFG = ε_uncond + scale × (ε_cond - ε_uncond)
• Scale controls prompt influence strength
• Randomly drop conditioning during training (10%) to learn unconditional generation
• Default scale is 7.5, too high causes oversaturation

---

课后练习 | Homework

1. CFG 实现：实现 CFG，对比不同 scale（1, 3, 7.5, 15）的输出差异
2. 丢弃训练：在训练时实现条件丢弃机制
3. 负向 Prompt：理解负向 prompt 的本质（也是 CFG）
4. 动态 CFG：实现随时间步变化的 CFG scale
5. 思考题：为什么 CFG 只在推理时用，训练时不用？