How does classifier-free guidance steer a diffusion model without training a classifier?

## The problem it solves A conditional diffusion model can be trained straightforwardly: feed the condition `c` alongside the noisy input and the step index, and the network learns `eps_theta(x_t, t, c)`. Left alone, such models often follow the condition loosely - the conditioning signal is one input among many, and nothing in the objective forces the samples to be strongly characteristic of `c`. The first fix was **classifier guidance**. Train a separate classifier that accepts *noisy* inputs at every noise level, take the gradient of `log p(c | x_t)` with respect to `x_t`, and add it to the score estimate the diffusion model provides. It works, but it demands a second model trained on the corrupted-input distribution at all noise levels - an awkward artefact to build, maintain and keep in sync with the generator, and one that only supports conditions expressible as classifier outputs. ## The construction Classifier-free guidance obtains the same steering from a single network by making it serve two roles. **Training.** Keep the ordinary conditional objective, but with some probability - commonly about ten percent - replace `c` with a learned null embedding standing for 'no condition'. Because the step index is drawn uniformly and the dropout is independent per example, the same weights end up estimating both the conditional noise and the unconditional noise across the whole chain. There is no second network, no second optimiser, no extra loss term. **Sampling.** At each reverse step, evaluate the network twice - once with `c`, once with the null token - and combine: ``` eps_hat = eps_uncond + w * (eps_cond - eps_uncond) ``` Equivalently `eps_hat = (1 - w) * eps_uncond + w * eps_cond`. Setting `w = 0` gives unconditional generation; `w = 1` gives ordinary conditional generation; `w > 1` extrapolates beyond the conditional prediction along the direction that distinguishes conditioned from unconditioned behaviour. The combined `eps_hat` is then fed into the usual reverse-step arithmetic in place of the raw prediction. ## Why the difference vector is the right direction Read through the score lens. The noise prediction is a rescaled score estimate, so `eps_cond - eps_uncond` is proportional to the difference between the score of the conditional noised density and the score of the unconditional one. By Bayes' rule that difference is proportional to the gradient of `log p(c | x_t)` - the very quantity classifier guidance obtained from an external classifier. The two conditional evaluations of one generator therefore recover an *implicit* classifier gradient without anyone ever training a classifier. That equivalence is the theoretical content of the method and the answer interviewers are listening for. ## Practical consequences - **Cost.** Two evaluations per reverse step. In practice the conditional and unconditional inputs are stacked into a single batch, so the wall-clock hit is closer to a wider batch than to a doubled step count, but the arithmetic is genuinely doubled. - **The null token is learned, not zero.** It is an embedding trained alongside everything else; zeroing the conditioning inputs instead usually produces an out-of-distribution input the network has never seen. - **Dropout is a training-time device only.** At sampling you deliberately evaluate both branches; you do not randomly drop the condition. - **The dropout rate is a trade.** Too low and the unconditional branch is undertrained, making the difference vector noisy; too high and conditional capacity is wasted on examples that carry no condition. - **It generalises past class labels.** Anything that can be embedded - a text encoding, a reference vector, a robot goal pose - can be dropped and nulled the same way, which is why the technique became the default conditioning mechanism across domains. ## Common misreadings Candidates often say the unconditional branch is 'a second model'. It is the same weights with a different conditioning input. Others claim the method needs no extra compute at sampling; it needs double. Others believe the conditioning dropout happens at generation time; it happens only during training. And some describe the mix as an interpolation between the two predictions - it is an interpolation only for `w` between zero and one, and the interesting regime is the extrapolation above one. ## A compact answer One network, conditioning replaced by a learned null token on a small share of training examples, two evaluations per reverse step, and predictions combined as unconditional plus `w` times the conditional-minus-unconditional difference - which is an implicit classifier gradient by Bayes' rule, obtained without ever training a classifier.