classifier guidance
**Classifier Guidance** is a technique for conditioning diffusion model generation on class labels or other attributes by using the gradients of a separately trained classifier to steer the sampling process toward desired classes. During reverse diffusion sampling, the classifier's gradient ∇_{x_t} log p(y|x_t) is added to the score function, biasing the generated samples toward inputs that the classifier confidently assigns to the target class y.
**Why Classifier Guidance Matters in AI/ML:**
Classifier guidance was the **first technique to achieve photorealistic conditional image generation** with diffusion models, demonstrating that external classifier gradients could dramatically improve sample quality and class fidelity without modifying the diffusion model itself.
• **Guided score** — The conditional score decomposes as: ∇_{x_t} log p(x_t|y) = ∇_{x_t} log p(x_t) + ∇_{x_t} log p(y|x_t); the first term is the unconditional diffusion model score, the second is the classifier gradient that pushes samples toward class y
• **Guidance scale** — A scalar parameter s controls the strength of classifier influence: ∇_{x_t} log p(x_t|y) ≈ ∇_{x_t} log p(x_t) + s·∇_{x_t} log p(y|x_t); larger s produces more class-specific but less diverse samples, with s=1 being standard Bayes and s>1 amplifying class fidelity
• **Noisy classifier training** — The classifier must operate on noisy intermediate states x_t at all noise levels, not just clean images; it is trained on noise-augmented data with the same noise schedule as the diffusion model
• **Quality-diversity tradeoff** — Increasing guidance scale s improves FID (sample quality) and classification accuracy up to a point, then degrades diversity and introduces artifacts; the optimal s balances sample quality against mode coverage
• **Limitations** — Requires training a separate noise-aware classifier for each conditioning attribute, doesn't generalize to text conditioning easily, and the classifier can introduce adversarial artifacts; these limitations motivated classifier-free guidance
| Guidance Scale (s) | FID | Diversity | Class Accuracy | Character |
|-------------------|-----|-----------|----------------|-----------|
| 0 (unconditional) | Higher | Maximum | Random | Diverse, unfocused |
| 1.0 (standard) | Moderate | Good | Moderate | Balanced |
| 2.0-5.0 | Lower (better) | Moderate | High | Sharp, class-specific |
| 10.0+ | Higher (worse) | Low | Very high | Oversaturated, artifacts |
**Classifier guidance pioneered conditional generation in diffusion models by demonstrating that external classifier gradients could steer the sampling process toward desired attributes, achieving the first photorealistic class-conditional image generation and establishing the gradient-guidance paradigm that inspired the more practical classifier-free guidance method used in all modern text-to-image systems.**