Classifier-Free Guidance
A technique to steer diffusion models toward a condition (like text) without training a separate classifier - by blending conditional and unconditional predictions.
Think of it like
Instead of hiring a critic to rank your sketches, you sketch two versions (one from imagination, one from description) and blend them.
Example
During training, sometimes drop the text prompt (unconditional). At generation, predict both paths - with and without text - then interpolate. This nudges the model toward the text without an external model.
How it actually works
Classic guidance requires a separate trained classifier (expensive, adds latency). CFG trains with random prompt dropout, so the model learns unconditional generation. At inference, compute both: unconditional pred and conditional pred, then blend: final = uncond + scale x (cond - uncond). Introduced in GLIDE; now standard in Stable Diffusion and others. No extra model, simple, effective. Scale controls influence; high scale can over-flatten diversity.
For product teams
Cheaper than training a classifier; enables real-time control without external dependencies.
For engineers
Requires training with conditional dropout. At inference, two forward passes. Typical scale: 7-15.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome