03.3 Generating an image through iterative denoising

01 | FROM A NOISE PATTERN TO A CANDIDATE IMAGE

A diffusion system can begin with a random pattern and gradually produce a coherent image. It does this through iterative denoising : repeated model-guided updates rather than one complete transformation.

The 2027 booklet identifies this repeated refinement as a strength of diffusion models and a source of computational demand. For Visionary Studios, understanding the process helps connect image quality with generation time and resources.

Begin with the central distinction: training has already taught the model useful patterns. Ordinary generation now uses those learned parameters to create a new sample.

02 | LOAD THE TRAINED SYSTEM

A usable generation system needs the trained model and its supporting components. Depending on the architecture, these may include a text encoder, a sampling procedure and an image decoder.

The model’s learned parameters provide the basis for its predictions. Loading them does not mean retraining them.

The booklet does not specify a particular software package for Visionary Studios. This lesson describes the process conceptually rather than giving instructions for one product.

03 | PREPARE ANY GUIDANCE

In conditional generation, the system processes guidance such as a text prompt. For our fictional campaign, that could describe an illustrated explorer beside a mountain lake.

The guidance influences the model’s predictions during sampling. It is not simply a caption attached after an unrelated image has been generated.

Unconditional generation omits specified content guidance for each sample but still uses a trained model and generation inputs. Not every diffusion system must receive text.

04 | SAMPLE THE STARTING NOISE

Generation from scratch commonly starts with a sample from a simple noise distribution, usually Gaussian noise in the diffusion setup considered here.

A seed can initialise the pseudorandom process. The noise has a numerical shape compatible with the system, such as an image-like array or a latent array.

It does not contain one predetermined photograph hidden under interference. Learned patterns, guidance and the sampling procedure shape what emerges.

05 | IDENTIFY THE CURRENT NOISE LEVEL

A sampler follows a sequence of noise levels or timesteps. In the usual reverse-process explanation, it moves from a highly noisy state towards a cleaner state.

The model receives information identifying the current level, because a prediction appropriate to strong corruption differs from one for a nearly clean sample.

Sampling timesteps need not be every timestep used to define training. Different samplers can choose different sequences, so a training timestep count is not automatically the number of model calls made for one image.

06 | MAKE A DENOISER PREDICTION

The trained denoiser processes the current sample and noise-level information, with any supported conditioning. In the common parameterisation introduced in 03.2, it predicts noise.

Other parameterisations predict clean data or a related quantity. The key idea is that the prediction provides information for the sampling update.

The system is not comparing the new request with a known correct final photograph. It is using a learned predictor on the current state.

07 | THE SAMPLER CALCULATES THE NEXT STATE

The sampler uses the model prediction and its defined update rule to calculate a next state at a lower noise level. Schedule-dependent coefficients matter.

Some sampling updates include fresh randomness; others are deterministic given their inputs and settings. Do not assume every update must add randomness or that every sampler behaves identically.

“Subtract the predicted noise” is a useful first intuition but not a complete DDPM algorithm. Actual updates account for the forward process and the selected sampling method.

08 | REPEAT ON THE UPDATED SAMPLE

The next prediction uses the updated state, not an unchanged copy of the original starting noise. Each update changes the input for a later step.

This dependency explains why ordinary sampling is iterative. Some internal calculations can be parallelised, but one sample’s next state commonly depends on the previous update.

09 | STEP THROUGH A SCHEMATIC

Select “Next stage” to follow a prepared sequence from a noise-like display to a simple mountain scene. Read the explanation at each stage.

This is a schematic illustration, not AI inference. The finished scene is hand-defined in the page, and display blends are precomputed. A real generator does not receive that finished target and reveal it. The illustration only helps you visualise repeated refinement.

10 | WHAT CAN CHANGE THROUGH THE PROCESS?

A sequence may develop recognisable structure and refine details as sampling proceeds. However, “first background, then objects, then texture” is not a universal fixed order.

Intermediate states can be difficult to interpret, and a visible preview may be an estimate of the eventual clean image rather than a direct display of the actual internal sample.

A realistic-looking intermediate image is not proof that sampling should stop immediately. Conversely, additional steps do not guarantee that every detail will improve.

11 | A SMALL UPDATE ANALOGY

Consider a deliberately simplified numerical illustration: the current value is 10 and a hypothetical prediction suggests a correction of 4. If our toy rule applies half of that correction, the next value is 10 − (0.5 × 4) = 8.

A subsequent step would use an updated value and a new prediction. It would not automatically subtract the same correction again.

This demonstrates dependence between steps, not an actual diffusion sampling equation. It has no trained model, schedule-derived coefficients or stochastic term.

12 | PIXEL SPACE AND LATENT SPACE

A pixel-space process works with image-value arrays. A latent diffusion process works with a learned representation and later decodes the result into an image.

In a latent system, the internal sample may have different dimensions and channels from the displayed RGB image. It should not automatically be described as a miniature photograph.

Working in a compressed representation can reduce some computational demands, but encoding and decoding introduce their own design choices. Not every diffusion model uses the same representation.

13 | DECODE AND PRESENT THE RESULT

After the final sampling update, a latent system decodes its final representation into image data. A pixel-space system has a different output path because sampling already operates in image space.

An application may then perform additional processing, such as resizing or safety checks. These are parts of the wider workflow rather than the definition of denoising itself.

The result is a candidate image. The studio still needs to check its subject, layout, character consistency, style and defects before approval.

14 | GENERATION STEPS ARE NOT LEARNING UPDATES

The trained network can be called repeatedly with the same learned parameters. Its activations and predictions change because the inputs change; its weights do not need to change.

An optimizer update during training changes those weights using a learning objective. A sampling update changes the generated sample using the trained predictor.

15 | STEPS, MODEL CALLS AND COMPUTATION

More sampling steps commonly involve more computation when the model and other conditions are comparable. But the number of steps alone does not fully describe generation cost.

Some guidance methods make multiple model evaluations per step. Resolution, model size, hardware, batch size and the sampler also affect time and memory.

For an invented classroom estimate, 20 model evaluations at 0.1 seconds each take 2 seconds of model-call time. This excludes text processing, decoding, communication and review, so it is not a complete workflow measurement.

16 | QUALITY AND REPRODUCIBILITY

Too few updates for a chosen model and sampler may give poor results, while additional updates can eventually offer little benefit or change the result in unhelpful ways. There is no universal ideal step count.

To investigate a result, record the model version, sampler, guidance, seed and relevant settings. Reproduction may still depend on software and hardware behaviour; a seed alone is not sufficient across arbitrary systems.

Compare repeated candidates against a fixed brief. One attractive output is weak evidence about consistent performance.

17 | APPLY IT TO VISIONARY STUDIOS

Visionary Studios needs to balance quality and computational efficiency. Repeated model use can create detailed outputs but can also increase waiting time and resource use.

The studio must also consider rejected candidates and editing. Finishing the sampling loop does not mean the campaign asset is ready to publish.

18 | EXPLAIN THE WHOLE GENERATION PATH

In your book, describe generation from the request to the reviewed output. Include guidance, starting noise, noise-level information, model predictions, sampling updates and any decoding.

Then explain two mistakes: “the final photograph is hidden inside the noise” and “every denoising step retrains the network”.

Read alongside the diffusion section of the official 2027 booklet. The scene illustration, toy update and timing figures are independent teaching examples.

CHECK YOUR UNDERSTANDING | PREDICT, SAMPLE, TRAIN OR DECODE?