Residual block with time embedding conditioning. §3.3 — "group normalization throughout... Transformer sinusoidal position embedding into eac

Self-attention block for the U-Net. §3.3 — "self-attention... at the 16×16 feature map resolution" Appendix B — "We add one head of self-atte

§3.2.2 — Multi-Head Attention. "Multi-head attention allows the model to jointly attend to information from different representation subspac

§3.3 — Position-wise Feed-Forward Networks. "FFN(x) = max(0, xW_1 + b_1)W_2 + b_2" (Eq. 2) "The dimensionality of input and output is d_mode

§3.3, Appendix B — U-Net noise prediction network ε_θ(x_t, t). "We use a U-Net backbone similar to an unmasked PixelCNN++ with group normali

U-Net configuration. Values from Appendix B (CIFAR-10 config) and official code unless marked.

§3.4, Eq. 14 — Simplified training objective L_simple. Computes MSE between true noise and predicted noise: L = ||ε − ε_θ(x_t, t)||²

§3.1 — Decoder: "The decoder is also composed of a stack of N = 6 identical layers."

§3.1 — Single decoder layer. "In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs

Spatial downsampling by factor 2. [FROM_OFFICIAL_CODE]

§4 — Exponential Moving Average of model parameters. "We report sample quality metrics using an exponential moving average (EMA) of model pa

§3.1 — Encoder: "The encoder is composed of a stack of N = 6 identical layers."

§3.1 — Single encoder layer. "Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple,

§5.4 — Label smoothing with KL divergence. "During training, we employed label smoothing of value ε_ls = 0.1. This hurt perplexity, as the m

All model hyperparameters. Values from {{PAPER_TITLE}} unless marked [UNSPECIFIED]. Matches configs/base.yaml — change values there, not

§3.5 — Positional Encoding (sinusoidal). "We use sine and cosine functions of different frequencies:" PE(pos, 2i) = sin(pos / 10000^(2i/d_m

Sinusoidal timestep embedding, following the Transformer positional encoding. §3.3 — "The diffusion time t is specified by adding the Transformer

§3 — The Transformer (Encoder-Decoder). "Most competitive neural sequence transduction models have an encoder-decoder structure. Here, the e

All model hyperparameters. Values from Vaswani et al. 2017, Table 3 ("base model") unless marked [UNSPECIFIED].

Spatial upsampling by factor 2. [FROM_OFFICIAL_CODE]

§5.1 — WMT 2014 English-German translation dataset. "We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 mill