Understand the technology
Give each name a clear job.
For May and November 2027. This introductory lesson is shared SL and HL content.
01 | TWO NAMES, TWO DIFFERENT ROLES
The case study refers to convolutional neural networks and the denoising diffusion probabilistic model. These names can sound like competing image generators, but they describe different levels of a system.
A CNN is a type of neural-network architecture. It can provide the learned prediction machinery inside a diffusion system. DDPM describes a probabilistic diffusion approach connecting a forward noising process, a learned reverse process, training and sampling.
A useful question is: “Which component makes a prediction, and which framework explains how that prediction is learned and used?”
02 | WHAT DOES CONVOLUTIONAL MEAN?
A convolutional layer applies a small numerical filter across positions in an input. At each position, it combines values in a local region to produce a response.
The same filter weights are reused across locations. This is called weight sharing . A filter can therefore respond to a relevant local pattern wherever it appears, rather than learning an entirely separate set of weights for each location.
Deep-learning libraries commonly implement this operation as cross-correlation, without reversing the filter. It is conventionally called convolution in CNN explanations. The worked example here uses that convention.
03 | PIXEL PATCHES AND FILTERS
A patch is a small region of the input. A kernel , or filter, contains weights that combine the patch values.
For a greyscale example, corresponding values are multiplied and the products are added. A bias may then be added. For inputs with multiple channels, the calculation also combines values across channels.
This is a local numerical operation. A filter does not receive an English label saying “this is a mountain”; it responds according to its weights and the supplied values.
04 | A WORKED FILTER CALCULATION
Use patch
[[1, 2], [3, 4]]
and filter
[[1, 0], [0, −1]]
, with zero bias. Multiply corresponding positions and add:
(1 × 1) + (2 × 0) + (3 × 0) + (4 × −1) = −3.
The response is one number for this filter at this position. Negative responses are possible. The small values are teaching numbers, not a complete image or a trained denoiser.
Moving the filter to other patches produces further responses. Arranging those responses by position creates a feature map.
05 | TRY THE LOCAL CALCULATION
Change the four patch values and calculate the filter response. The fixed filter is
[[1, 0], [0, −1]]
and the bias is zero.
Observe that only the top-left and bottom-right values affect this particular response because the other weights are zero. Other filters would combine the patch differently. This calculator performs one numerical operation, not neural-network training.
06 | WHAT IS A FEATURE MAP?
A feature map records responses across positions for a learned feature channel. Different filters can produce different channels, responding to different patterns.
Early layers can learn local relationships such as contrasts or textures. Later layers combine previous responses into more complex representations. Not every channel has a simple human-readable meaning.
A CNN does not need a separate hand-written rule for every visible feature. During learning, filter weights are adjusted through the training objective.
07 | NON-LINEARITY AND MULTIPLE LAYERS
Convolutions alone are linear operations. Networks also use non-linear functions and other operations, allowing them to represent more complex relationships than a sequence of purely linear transformations.
An activation is a computed value within the network. An activation function transforms values, often non-linearly. Activations depend on the current input; weights are learned parameters shared across inputs.
Changing a noisy sample changes activations even when the weights stay fixed. That helps explain how one trained network can produce different predictions at successive generation steps.
08 | LOCAL FILTERS CAN USE WIDER CONTEXT
The receptive field of a response is the input region that can influence it. Combining layers and using operations that change spatial resolution can let later responses depend on larger parts of the input.
Context matters for denoising. A dark value might be part of a genuine object boundary or a disturbance; neighbouring and wider patterns help the model estimate what is plausible.
Local filtering does not imply that the complete network can only consider one tiny patch. Its effective context depends on the whole architecture.
09 | A CNN AS A DENOISER
A convolutional denoiser can take a noisy sample, with timestep information and any supported conditioning, and output an estimate used by the diffusion process.
In our common noise-prediction example, the output estimates sampled noise. Other parameterisations estimate clean data or related quantities.
The CNN supplies a prediction. It does not, by itself, specify the complete forward noise schedule, learning objective or generation update rule.
10 | AN ARCHITECTURE EXAMPLE: U-NET
Many diffusion denoisers have used a U-Net -style architecture. A lower-resolution path builds representations with broader context, while a higher-resolution path builds an output with spatial detail.
Skip connections pass information between parts of the network, helping preserve useful spatial information. A diagram shaped like a U describes information flow, not the path followed by a physical image.
This is supporting context, not a claim that every diffusion system uses the same U-Net. Denoisers can include attention or use other architectures, including transformers.
11 | WHAT DDPM DEFINES
Denoising diffusion probabilistic model names a diffusion approach in which a specified forward process gradually noises data and a learned reverse process models movement back towards data-like samples.
The framework connects the corruption process to a training objective and a probabilistic generation procedure. A denoising network is a component within that arrangement.
“Probabilistic” means that distributions and uncertainty are part of the model. It does not mean the system makes arbitrary changes without learned relationships.
12 | FORWARD PROCESS: A KNOWN CORRUPTION
In the standard DDPM setup, the forward process introduces Gaussian noise according to a schedule. The clean contribution weakens while noise becomes more prominent.
The training process can construct a noisy sample at a selected level and retain knowledge of the sampled noise. That supports a learning task without requiring a person to annotate the noise manually.
The CNN normally receives the noisy sample and level information, rather than the exact target noise as an answer to copy.
13 | LEARNING: CONNECT INPUT AND TARGET
For a common DDPM parameterisation, the network predicts noise. A loss compares that estimate with the noise used to create the training example.
Backpropagation computes parameter gradients, and an optimizer adjusts learned weights, including convolutional filters where used. This repeats across examples and noise levels.
The learning objective connects the network to the diffusion task. A generic CNN trained to classify photographs would not automatically become a diffusion denoiser merely because it contains convolutions.
14 | REVERSE PROCESS: USE THE PREDICTION
During generation, a starting sample is drawn from a noise distribution. The trained predictor supplies information for reverse-process updates.
In the standard DDPM sampling formulation, reverse transitions are probabilistic. Other sampling procedures can use diffusion predictions differently, including deterministic updates in suitable formulations.
The sampler applies its update rule and schedule. It is not enough to say “the CNN subtracts all the noise once”. Generation usually calls the predictor repeatedly with fixed learned parameters.
15 | PUT THE ROLES SIDE BY SIDE
| Question | CNN | DDPM framework |
|---|---|---|
| What does the name describe? | A neural-network architecture using convolutions | A probabilistic diffusion learning and generation approach |
| Typical role here | Predict noise or another relevant quantity | Connect forward noising, learning and reverse sampling |
| Can it appear elsewhere? | Yes, for example in classification | It can use a denoising architecture as a component |
| Does it guarantee final image quality? | No | No |
They fit together rather than represent a simple choice between two equivalent products.
16 | A COMPLETE SYSTEM HAS MORE COMPONENTS
Prepare noisy data, timestep information and any conditioning.
A trained denoiser, potentially a CNN, estimates a relevant quantity.
Training uses a loss and optimizer; generation uses sampling updates.
A deployed text-guided latent system may additionally use a text encoder and image decoder. Not all these components are CNNs, and not every component is trained together in the same way.
Explain the component needed for the question rather than using “CNN” as a name for the entire application.
17 | APPLY IT TO VISIONARY STUDIOS
Understanding architecture helps explain how spatial patterns can be processed. Understanding DDPM helps explain why training and repeated generation require computation.
Output quality, consistency, resource use and data concerns still need evidence. The case study does not supply a winning architecture for the company.
18 | EXPLAIN THE TWO ROLES
In your book, write two short explanations: “What does the CNN do?” and “What does DDPM describe?” Connect them with a training example and a generation example.
Then calculate the response of the worked filter and explain why that one local calculation is not a complete denoising process.
Reveal a possible explanation
A CNN uses learned convolutions and other operations to compute predictions from an input. In a diffusion setup, it may predict noise from a noisy sample and timestep. DDPM connects forward noising with learning and a probabilistic reverse process.
Training compares prediction and target and updates parameters. Generation uses predictions in sampling updates with parameters normally fixed. The worked filter gives −3, one local response; a real predictor combines many learned operations across a much larger representation.
Read alongside the diffusion section of the official 2027 booklet. The numerical filter and architecture explanations are independent teaching extensions.
CHECK YOUR UNDERSTANDING | ARCHITECTURE OR DIFFUSION FRAMEWORK?