☆ Save Reverse Diffusion Process — How Diffusion Models Generate Data from Noise

09/13/2026

The reverse diffusion process generates new data by starting from random noise and refining it step by step. Let xt represent the current state at time step t. The model predicts the noise present in xt and uses that estimate to produce a slightly less noisy state xt-1. Repeating this transition over many time steps gradually turns the initial noise xT into a structured sample x0. The key idea is not to predict the finished result all at once, but to improve the current state through a sequence of small reverse transitions.

In simple terms: For image generation, xT starts as random noise with no recognizable structure. A denoising network estimates the noise in the current state, and the reverse transition uses that prediction to produce the next state in the reverse process, xt-1. That new state is fed back into the model, and the same procedure is repeated. As the time index moves toward zero, recognizable image structure gradually emerges. The final x0 is not a reconstruction of one particular training image; it is a newly generated sample drawn from the data distribution learned by the model.

Reverse diffusion gradually transforms random noise into newly generated data

How It Works

• The Current State xt and the Direction of Generation

  • xt represents the current data state at a particular time step t. In image generation, it is not a single number but a tensor containing many values that together encode the current image state.
  • When t is large, early in generation, the state is dominated by noise. As t decreases, more recognizable data structure appears. The process begins at xT, which is close to pure random noise.
  • The model does not convert xT directly into a finished image. Instead, it takes the current xt, produces a slightly less noisy xt-1, and then uses that result as the input to the next reverse step.
  • The reverse trajectory is therefore xT → xT-1 → … → xt → xt-1 → … → x0. Although xt-1 is the previous state in terms of the time index, it is the next state generated by the reverse process. A small matrix example later in this section shows how one such transition changes the actual values.

• Denoising Network and Noise Prediction

  • To move backward through the diffusion trajectory, the model needs an estimate of the noise contained in the current state xt. That estimate comes from the denoising network. Given xt and the time step t, the neural network predicts the noise present in the current state.
  • The predicted noise is written as εθ(xt, t), where θ denotes the learned parameters of the denoising network. If xt is a tensor representing the current image state, εθ(xt, t) has the same shape. Each entry estimates the noise associated with the corresponding position in xt.
  • During training, noisy states xt are created by adding known noise to original samples at different time steps. Because the training procedure knows exactly which noise was added, the denoising network can learn by comparing its prediction with that known target. After training, the model can estimate the noise in xt without access to the original clean sample.
  • The predicted noise then provides the information needed to estimate xt-1. Reverse diffusion is therefore not the forward process executed backward as an exact inverse. It is a generative procedure that uses learned noise prediction to estimate a distribution for the previous-time state and sample from it.

• Probabilistic Computation of the Reverse Transition

  • Once the denoising network predicts the noise, the model uses that estimate to move from xt to xt-1. In a DDPM, this one-step update is the reverse transition. Probabilistically, the model defines a distribution for xt-1 conditioned on the current xt, then samples the actual state used in the next reverse step from that distribution.

  • The calculation also uses quantities that describe the noise schedule. βt specifies the predefined noise level at time step t, while αt=1−βt describes how much of the existing signal is retained in that step. ᾱt is the cumulative product of the α values from the first step through t, written as ᾱt=∏s=1tαs. It summarizes how far xt has progressed through the overall noising process.

  • $$\mu_\theta(x_t,t)=\frac{1}{\sqrt{\alpha_t}}\left(x_t-\frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}_t}}\epsilon_\theta(x_t,t)\right),\qquad x_{t-1}=\mu_\theta(x_t,t)+\sigma_t z$$

    The first expression computes the center μθ(xt,t) of the reverse transition, and the second samples xt-1 by adding the stochastic term σtz.

  • The equation is easier to read when separated conceptually into two stages. First, xt and the predicted noise εθ(xt, t) are combined to compute the center of the distribution from which xt-1 will be sampled. The terms αt and ᾱt scale that calculation according to the noise level associated with the current time step.

  • The second stage adds σtz, which introduces stochastic variation around the computed center. The model therefore does not map xt to one fixed xt-1. Instead, it follows the sequence predict the noise → compute the center of the next reverse-state distribution → sample an actual xt-1 around that center. The next example shows how the same calculation works when the state is represented as a small tensor.

• Following a Reverse Transition with a Small Matrix

  • In a real image model, xt is a large tensor, so looking at every value at once makes the calculation difficult to follow. To keep the mechanics visible, we will use the same simplified 3×3 matrix throughout the example. These values are only an illustrative tensor representation and should not be interpreted specifically as RGB pixel values.

  • $$x_t=\begin{bmatrix}0.72&-0.31&0.45\\0.18&0.91&-0.52\\-0.24&0.36&0.67\end{bmatrix}$$

    A simplified 3×3 representation of the current image state xt at time step t.

  • The denoising network predicts a noise tensor rather than a single scalar. Suppose it receives the current xt and produces the following εθ(xt, t). Each entry represents the direction and magnitude of the noise the model estimates at the corresponding position.

  • $$\epsilon_\theta(x_t,t)=\begin{bmatrix}0.12&-0.08&0.20\\0.04&0.15&-0.11\\-0.06&0.09&0.13\end{bmatrix}$$

    The noise predicted by the denoising network has the same simplified 3×3 shape as xt.

  • The first stage of the reverse transition combines xt with the predicted noise and scales the result using αt and ᾱt. Suppose this calculation gives the following center for the xt-1 distribution. The important point is that the result is not a single scalar such as 0.60. It is another matrix with the same shape as xt.

  • $$\mu_t=\begin{bmatrix}0.68&-0.27&0.36\\0.16&0.84&-0.47\\-0.21&0.32&0.61\end{bmatrix}$$

    A simplified center matrix μt for xt-1, computed after accounting for the predicted noise.

• σtz and Stochastic Sampling

  • Next comes the σtz term from the reverse transition. Here, z is also not a single number. It is a newly sampled random-noise tensor with the same shape as the current state. Suppose the following z is drawn for this time step.

  • $$z=\begin{bmatrix}0.20&-0.10&0.30\\-0.20&0.10&0.00\\0.10&-0.30&0.20\end{bmatrix}$$

    A simplified random-noise sample used to introduce stochastic variation in this reverse transition.

  • σt controls how strongly this random variation affects the current transition. For illustration, let σt=0.1. Multiplying each entry of z by 0.1 reduces the magnitude of the stochastic adjustment before it is added to the center.

  • $$\sigma_tz=\begin{bmatrix}0.02&-0.01&0.03\\-0.02&0.01&0.00\\0.01&-0.03&0.02\end{bmatrix}$$

    With σt=0.1, the sampled noise is scaled to produce the stochastic variation used in this step.

  • Finally, σtz is added element by element to the center matrix μt. This produces the actual xt-1 sampled for the current reverse transition.

  • $$x_{t-1}=\begin{bmatrix}0.68&-0.27&0.36\\0.16&0.84&-0.47\\-0.21&0.32&0.61\end{bmatrix}+\begin{bmatrix}0.02&-0.01&0.03\\-0.02&0.01&0.00\\0.01&-0.03&0.02\end{bmatrix}$$

    $$x_{t-1}=\begin{bmatrix}0.70&-0.28&0.39\\0.14&0.85&-0.47\\-0.20&0.29&0.63\end{bmatrix}$$

    The sampled state xt-1 is obtained by adding a small stochastic variation to the center matrix.

  • The σtz term does not simply add back the noise that was previously removed. Its role is to introduce stochastic variation so that the model can sample one specific state around the center of the reverse distribution. The resulting xt-1 becomes the current state for the following reverse step, where the denoising network again predicts a full noise tensor.

• How Repeated Matrix Updates Become an Image

  • The xt-1 produced above is only one intermediate state. In the following reverse step, that entire tensor becomes the new input, and the denoising network once again predicts a noise tensor with the same shape.
  • The new prediction is used to compute another center, and adding the corresponding σt-1z term produces xt-2. Throughout reverse diffusion, the tensor keeps the same overall shape while its values change from one time step to the next.
  • Beginning with xT, the model repeats these updates until it reaches x0. A state dominated by random noise gradually develops large-scale structure and then finer details. The model is not generating a completely new image at every step; each reverse transition incrementally updates the state produced by the previous one.
  • The overall image-generation trajectory can therefore be summarized as current tensor xt → predict a noise tensor → compute the center tensor → add stochastic variation → sample xt-1. The sampled xt-1 becomes the input to the following step, and the same process continues until x0 is reached.

Significance and Limitations

The main idea behind the reverse diffusion process is to avoid solving the entire generation problem in one step. Instead, the model repeatedly solves a smaller problem: given the current state, estimate and sample the previous-time state in the reverse trajectory. The denoising network predicts the noise in the current tensor, and the reverse transition uses that prediction to sample the next state in the generation sequence. As these updates accumulate across many time steps, the initial random noise is transformed into a new sample that follows the learned data distribution.

The tradeoff is speed and computational cost. Generating a single x0 requires repeated denoising-network evaluations and reverse transitions across many sampling steps. Using fewer sampling steps can make generation faster and reduce computation, but reducing them too aggressively can hurt the quality or stability of the result. Practical diffusion systems therefore aim to preserve the reverse diffusion mechanism while reaching high-quality samples with as few steps as possible.

Recommended prerequisite reading (3/5)

+2

• Reverse Distribution — Estimating the Previous State in Diffusion Models
• Noise Scale — How Noise Levels Shape Denoising Quality in Diffusion Models
• Noise-Conditioned Score Network — Why Diffusion Models Learn the Score Function

• Generative AI — A Model That Learns Data Distributions to Create New Samples
• Text-to-Video Generation — Video Generation Conditioned on Text Prompts

Recommended next reading (5/15)

+5

• 8. Generative Modeling — Core Principles of Generative Models and Deep Generative Models
• Noise Corruption — How Denoising and Self-Supervised Learning Learn from Noisy Inputs
• PixelRNN — How Autoregressive Models Generate Images One Pixel at a Time
• Text-to-Image Models — Multimodal Generative Models That Synthesize Visual Images from Textual Meaning
• Invertible Transformations — How Normalizing Flows Preserve Information

• Identity Preservation — Keeping the Same Identity Consistent Throughout the Generation Process
• 7.3 Self-Attention Mechanism — From Query–Key–Value to Matrix Computation
• Score Matching — Learning Data Distributions Through the Score Function
• 7.8 Advanced Positional Embeddings — APE·RPE·and RoPE in Transformer Models
• Score-based Generative Model — Learning the Generative Direction with the Score Function

• Normalizing Flow — A Generative Model That Learns Probability Density Through Invertible Transformations
• Restricted Boltzmann Machine (RBM) — An Efficient Generative Model That Learns a Probability Distribution with Two Connected Layers
• Score Function — Understanding the Gradient of Log Probability Density
• Implicit Density — Representing a Distribution Without Explicitly Computing Its Density
• Stochastic Encoder — Learning Latent Representations Through Probability Distributions

Posts on the same topic (2/2)

• Music Generation — A Generative Approach That Builds a Full Piece by Predicting and Extending the Next Sound Over Time
• Latent Video Representation — Compressing Spatial and Temporal Information in Video

Related concepts (2/2)

• Denoising Diffusion Probabilistic Model (DDPM) — Generating Data by Predicting Noise
• Diffusion Model — How Denoising Turns Random Noise into New Data

📍 Where this concept fits in the AI learning map

See where this concept sits within the full AI Universe.

📍 Current position in AI Universe

☰

Reset Show completed · Login required Loading…

🌌 AI Universe

‹

›

⭐ Concept

Select a star.

View the full AI Universe

« Noise Corruption — How D…|ARIMA — A Linear Time Se… »

🔖 Tags: Denoising Network · diffusion-model · Generative Model · Noise Prediction · Reverse Diffusion Process · Reverse Transition · sampling