Conditioning paths: identify where time and content enter the denoiser

A timestep input may be an integer schedule index, a continuous noise coordinate, or a transformed quantity such as log-SNR. Its embedding must match the model's training convention. Two schedules with the same integer index can represent different corruption levels. External conditions can include class labels, text tokens, spatial masks, or other structured inputs, each with its own encoder and alignment. Do not substitute a scalar class ID for a text embedding or change time units while preserving only the tensor shape.

A spatial condition can be concatenated to image or latent channels, while token sequences can enter through cross-attention. Some architectures combine several mechanisms. Preserve batch correspondence, image-coordinate transforms, padding masks, and null representations after augmentation or cached preprocessing. Freezing an encoder's parameters does not automatically fix its dropout mode or tokenizer revision. A model can silently ignore a malformed condition and still reduce denoising loss, so inspect intermediate shapes and controlled condition changes rather than relying only on successful execution.

A condition embedding can produce featurewise scale and shift values for normalized hidden features. A separate coefficient can gate a residual branch. DiT studies conditioning choices including adaptive layer normalization and an initialization variant that starts selected modulation or residual outputs at zero. Those are concrete architecture and initialization decisions, not a claim that every diffusion transformer uses the same block. A learned scale or residual gate need not be a probability or lie between zero and one.

Cross-attention and adaptive normalization also expose different capacity and cost tradeoffs. Text length changes attention work, whereas a pooled condition vector can discard token-level distinctions. Spatial concatenation preserves another alignment structure. Compare these paths with matched architecture and training budgets where possible, and report their actual conditioning encoders. A larger condition representation can improve one task while increasing memory, introducing leakage from an unintended input, or making the null branch inconsistent with inference.

Hold noisy input and time fixed while changing only the condition, then inspect outputs and intermediate activations. Shuffle conditions as a diagnostic, compare real and null paths, and verify gradients reach the intended trainable modules. A response difference establishes sensitivity, not correct semantic alignment or improved sample quality. Check masking and spatial transforms on hand-labeled examples. For checkpoint interchange, verify encoder outputs, normalization, patch/grid geometry, time convention, and target parameterization together rather than matching only parameter names.

The network lab shows one bounded scalar gate. Reset, set x=1 and w=1, predict σ(0.5)≈0.622459, then Play or Step. Separately Reset, set x=1 and w=2, predict σ(1.5)≈0.817574, then Play or Step. It has no timestep embedding, attention, normalization, or training. Its sigmoid range should not be transferred to the unconstrained coefficients in the teaching residual block. Use the instrument only to inspect one explicit signal path before deriving the actual backbone's transformations.