Conditioning dropout: train a null branch before choosing a guidance scale

In the teaching setup, independently replace the complete external condition with a designated null representation with probability p_drop. Keep the noisy input, timestep, target, and clean-data population otherwise unchanged. A learned null class embedding, an encoded empty prompt, and a zero tensor are different representations; use the one the model was trained to understand. Dropping individual text tokens or leaving another condition channel visible does not necessarily train the same null branch.

With a constant condition-independent dropout probability, the null branch sees the corresponding marginal training-data mixture. If dropout depends on class or sample properties, that mixture can change. Conditioning masks can be per example or follow another declared grouping, which alters finite-batch behavior. Record the random-generator state and mask construction. The expected number of null examples in a batch is not a promise that every batch contains one, and a tiny dropout rate can leave that branch poorly trained under a limited budget.

Use the convention epsilon_guided=epsilon_null+w(epsilon_cond−epsilon_null). Here w=0 selects the null prediction, w=1 selects the conditional prediction, and w>1 extrapolates beyond it. Some papers write conditional plus another coefficient times conditional-minus-null, shifting the numerical scale by one. Record the actual formula instead of comparing a number called guidance scale across implementations. Both predictions must use the same noisy sample, time, and compatible target parameterization.

Guidance changes the vector field used by the sampler and commonly trades conditional adherence against other quality or diversity properties. It is not the same operation as increasing training dropout or multiplying the training loss. Even with ideal score estimates, a pointwise guided score expression does not by itself prove a numerical sampler draws exactly from a desired global powered posterior. Finite steps, model errors, clipping, rescaling, and negative-prompt substitutions add further assumptions. Evaluate the actual inference configuration rather than treating guidance as a guarantee.

To study dropout, keep architecture, target, data budget, and an evaluation guidance protocol fixed. To study guidance, freeze the checkpoint and vary only the declared inference scale and sampler controls. Measure null and conditional branch behavior, conditional alignment, sample quality, diversity, and failures across prompt groups. Two forward branches can be batched together, but they still require computation and memory under the chosen implementation. A better-looking selected grid does not replace a fixed prompt and seed portfolio with quantitative and human evaluation.

The probability lab models independent null-mask counts. Reset, set p=0.10 and n=20, predict mean two null examples and zero-null probability about 0.1216, then Play or Step. Separately Reset, set p=0.20 and n=20, predict mean four and zero-null probability about 0.01153, then Play or Step. The exact binomial has no denoiser, text encoder, or inference guidance. It illustrates why training exposure is a distribution, while the hand calculation determines the separate extrapolated prediction.