Guidance Scales in Diffusion Models

Module: Generative AI | Difficulty: Advanced

Guidance Scale Theory

Effective Temperature

Higher = lower temperature = sharper attention.

Dynamic Guidance

Schedule guidance strength over time.

Negative Prompt Guidance

def dynamic_guidance(model, x_t, t, cond, uncond, w_min=3, w_max=12):
    progress = t / 1000
    w = w_min + (w_max - w_min) * progress
    eps_cond = model(x_t, t, cond)
    eps_uncond = model(x_t, t, uncond)
    return eps_uncond + w * (eps_cond - eps_uncond)

Research Insight: The optimal guidance scale depends on the trade-off between quality and diversity. For text-to-image, w=7-12 is optimal. For class-conditional, w=1-3 is optimal. Dynamic scheduling can improve results by 5-10%.