AIFEATURE

How Diffusion Models Turn Random Noise Into Images: Add Noise, Then Learn to Remove It

N
NathanTechnology Editor · Technical Lead
Published · Updated
Diffusion probabilistic models generate images by running two linked processes: a forward process that gradually buries an image in Gaussian noise, and a reverse process trained to predict that noise and denoise pure randomness back into a clean image. The original DDPM formulation reached an Inception score of 9.46 and an FID of 3.17 on CIFAR10, and the same principle now powers production tools like Stable Diffusion.

What Is a Diffusion Probabilistic Model?

The original Denoising Diffusion Probabilistic Models (DDPM) paper presents diffusion probabilistic models as a class of latent variable models, inspired by nonequilibrium thermodynamics, that produce high-quality image synthesis resultsCITE:E1. That framing — a physics-inspired latent variable model rather than a single fixed generation step — is the foundation the rest of the technique builds on: an image is not produced in one shot, but recovered gradually across many steps.

How Does the Forward Process Gradually Corrupt Images With Noise?

The forward process gradually destroys an image by adding Gaussian noise at each timestep and then rescaling the result to keep its size constantCITE:E7CITE:E3. NVIDIA frames this as the "easy direction" to start from: begin by considering how to destroy an image by gradually adding noise onto itCITE:E7. At each timestep t, a small amount of Gaussian noise is added to the previous image, and the image is then rescaled to maintain a constant image sizeCITE:E3. Repeated enough times, this sequence turns a structured image into what is effectively pure Gaussian noise.

How Does the Model Learn to Predict and Remove Noise in Reverse?

The reverse process trains the model to undo each noising step, denoising pure Gaussian noise into a clean imageCITE:E4. In learning to perform the reverse of the noising steps of the forward process, the model is essentially learning to denoise pure Gaussian noise into a clean imageCITE:E4. Critically, at each image step the model is trained to predict the entire noise that was added, rather than the specific amount of noise to be removed at that particular stepCITE:E5. That distinction — predicting the full noise rather than a single step's residual — is the specific mechanism that lets the reverse process reconstruct an image from randomness.

What Results and Applications Have Diffusion Models Achieved?

DDPM reached an Inception score of 9.46 and a then-state-of-the-art FID score of 3.17 on the unconditional CIFAR10 datasetCITE:E2. On the same unconditional CIFAR10 benchmark, the paper reports both figures together as evidence of image qualityCITE:E2. The same denoising principle extends beyond benchmarks into applied products: Stable Diffusion is a generative AI model that produces unique photorealistic images from text and image promptsCITE:E6.

Taken together, the evidence traces one mechanism end to end. The forward process buries an image in Gaussian noise step by stepCITE:E7CITE:E3, and the reverse process is trained specifically to predict the entire noise added at each step rather than a stepwise residual, so it can denoise pure Gaussian noise back into a clean imageCITE:E4CITE:E5. That is the same latent-variable formulation the original DDPM paper describesCITE:E1, backed by concrete CIFAR10 scores of Inception 9.46 and FID 3.17CITE:E2, and it is the principle underlying Stable Diffusion's text-and-image-prompt image generationCITE:E6.

📊 Evidence

FAQ

What Is a Diffusion Probabilistic Model?

The original Denoising Diffusion Probabilistic Models (DDPM) paper presents diffusion probabilistic models as a class of latent variable models, inspired by non…

How Does the Forward Process Gradually Corrupt Images With Noise?

The forward process gradually destroys an image by adding Gaussian noise at each timestep and then rescaling the result to keep its size constantCITE:E7CITE:E3.

How Does the Model Learn to Predict and Remove Noise in Reverse?

The reverse process trains the model to undo each noising step, denoising pure Gaussian noise into a clean imageCITE:E4.

What Results and Applications Have Diffusion Models Achieved?

DDPM reached an Inception score of 9.46 and a then-state-of-the-art FID score of 3.17 on the unconditional CIFAR10 datasetCITE:E2.

📎 Sources

  1. arxiv.org
  2. ibm.com
  3. aws.amazon.com
  4. developer.nvidia.com

Related data

Author's TakeNathan

The detail that matters most in this evidence set is the specific training target: the model is trained to predict the entire noise added at each image step, not the smaller residual amount to remove at that step. That single design choice is what connects the forward process's simple, repeated noise addition to a reverse process capable of reconstructing an image from pure Gaussian noise. The CIFAR10 numbers give that mechanism a concrete anchor — Inception 9.46 and FID 3.17 — and Stable Diffusion's move to generating photorealistic images from text and image prompts shows the same denoising principle deployed as a production tool rather than a benchmark exercise. The metric worth watching next is whether later diffusion work reports FID scores below the 3.17 baseline DDPM established, since that number is the one concrete quality yardstick this evidence set provides.

N
NathanTechnology Editor · Technical Lead

Related

BRIEF

GreenTrans Unveils GT5X, GT3X Quadruped Robots, Targets 100% Taiwan-Made Content by 2027

GreenTrans (綠捷), the robotics subsidiary of China Motor (中華車), unveiled quadruped robots GT5X and GT3X at SEMICON Taiwan 2026, targeting 100% Taiwan-made content by 2027. The robots combine an in-house-designed control unit and battery management system, NVIDIA's Jetson Orin and Isaac Lab platforms, and a new LFP battery developed with Formosa Smart Energy (台塑新智能), while GreenTrans's inspection robots are already deployed in semiconductor fabs.

EffectStory 編輯部 ·
BRIEF

Nvidia Confirms $12.93 Billion Acquisition of Hugging Face

Nvidia confirmed on September 3, 2026 that it agreed to buy Hugging Face for $12.93 billion, exactly $12,930,300,000, gaining the open-source AI hosting platform used by over 18 million developers. CEO Jensen Huang pledged the platform will stay open, with no Nvidia compute required to build on or deploy through it.

EffectStory 編輯部 ·
BRIEF

NVIDIA to Subscribe US$3.5 Billion of MediaTek's Record US$3.9 Billion Convertible Bond

NVIDIA will subscribe US$3.5 billion of MediaTek's US$3.9 billion offshore convertible bond offering, the largest such issuance in Taiwan's capital market history, deepening cooperation in AI infrastructure, edge AI computing, and automotive platforms while marking NVIDIA's first major investment in a Taiwanese company.

EffectStory 編輯部 ·