Understand the technology
From open-ended samples to useful data.
For May and November 2027. This introductory lesson is shared SL and HL content.
01 | GENERATE WITHOUT A CONTENT BRIEF
Imagine asking a trained image model to produce a sample without specifying a subject, class label, sketch or scene map. The model generates according to patterns learned during training rather than following a content brief for that request.
Unconditional generation produces samples without supplied content guidance for each output. The model still needs a generation process and may use random starting values or noise.
The 2027 booklet connects this approach with exploratory creativity, abstract art and synthetic datasets. For Visionary Studios, the key question is whether open-ended variation serves the task better than tightly specified guidance.
02 | UNCONDITIONAL DOES NOT MEAN NO INPUTS
A model cannot run without computation. It may take a random noise vector or an initial noise image and use learned parameters to produce a sample. These are generation inputs even when no content prompt is supplied.
The distinction concerns content conditioning : information that specifies what the requested sample should contain. A seed that initialises randomness is different from a label requesting “landscape”.
Changing a random starting value can change the sample without giving the model a semantic instruction about the desired subject.
03 | THE TRAINING DATA STILL SHAPES THE RESULT
Unconditional generation does not mean anything whatsoever is equally likely. The model has learned from a particular training dataset and learning process.
A model trained on animal images may generate animal-like samples. That reflects its learned distribution, even if the user supplies no animal label during generation.
A restricted training dataset and a class label supplied for each request are different forms of restriction. Selecting a specialised trained model is also not the same as passing a category condition into that model.
04 | SAMPLING FROM A LEARNED DISTRIBUTION
A distribution describes how values or examples are spread and how likely different outcomes are. A generative model aims to learn patterns that allow it to produce samples related to its training data.
Sampling means drawing an output through the model’s generation process. Different random starting values may lead to different samples.
For images, the learned relationships are complex: shapes, textures and arrangements can occur together. You do not need probability equations here, but you should understand that generation is shaped by learned patterns rather than an unrestricted imagination.
05 | CONDITIONAL AND UNCONDITIONAL COMPARED
| Question | Conditional generation | Unconditional generation |
|---|---|---|
| Is content guidance supplied for the sample? | Yes: for example text, a label or an image | No specified content guidance |
| Are learned parameters used? | Yes | Yes |
| Can randomness contribute? | Yes | Yes |
| Typical creative use | Working towards a particular brief | Exploring samples without specifying their content |
Removing the text from an interface does not prove the model is operating unconditionally. An application may still supply default prompts, reference images or other conditions behind the scenes.
06 | OPEN-ENDED CREATIVE EXPLORATION
Suppose Visionary Studios wants unexpected visual starting points before a campaign direction is chosen. Unconditional samples could offer patterns or scenes that designers had not explicitly requested.
A designer could select an interesting sample and develop it further through editing or a later conditional workflow. The role of the sample is inspiration rather than automatic approval as a final asset.
Unconditional generation does not necessarily produce abstract art. An unconditional model can produce recognisable objects if that is what it learned. Abstract output depends on the model and its training, not simply the absence of a prompt.
07 | WHEN OPEN-ENDED OUTPUT IS A POOR FIT
A client requiring a particular product, recurring character and layout gives the studio a specific target. Unconditional sampling does not directly communicate those requirements.
The team might reject many samples while searching for something usable. This can reduce the apparent advantage of fast generation.
For tightly controlled advertising, conditional guidance may be more appropriate. A recommendation should connect the approach to the stage of the creative process rather than claim that either approach is always superior.
08 | WHAT IS SYNTHETIC DATA?
Synthetic data is artificially produced data used for a purpose such as development, testing or training. In this lesson, we focus on generated images used as examples for another learning task.
For instance, a team might create artificial images of objects to help develop a recognition system. Those images are data for the downstream system, even though they are outputs of the generator.
Synthetic data is broader than unconditional generation. It can come from conditional models, simulations or other procedures. The terms describe different things: how an image is generated and how the resulting data is used.
09 | FOLLOW THE TWO-MODEL WORKFLOW
A trained system produces artificial images.
People or supported tools check, select and label images.
Another learning system uses the dataset for its task.
The generator’s training and the downstream model’s training are separate processes. Producing a batch of synthetic images is normally inference for the generator, even if those images later become training data elsewhere.
For example, generated animal images could be considered for a classifier. Their suitability would depend on their accuracy, variety and relevance to the classifier’s intended real-world inputs.
10 | GENERATED IMAGES DO NOT AUTOMATICALLY HAVE CORRECT LABELS
An unconditional generator may produce varied images without providing reliable category labels. Before supervised training, the dataset needs labels appropriate to the task.
Even a conditional request is not proof of the resulting content. If a system is asked for a dog but produces a distorted or ambiguous animal, automatically labelling the image “dog” can introduce errors.
Label verification checks whether the assigned label matches the actual image. Human review or other checks can help, but their reliability must also be evaluated.
11 | WHY USE SYNTHETIC IMAGES?
Artificial examples may help expand a dataset, explore variations or test situations that are difficult to collect. A controlled generator or simulator may allow particular viewpoints, backgrounds or lighting conditions.
These are potential benefits, not automatic outcomes. An unconditional model may not reliably produce a rare case merely because the dataset needs it; targeted conditions or simulation may offer more control.
Collecting and checking synthetic data also has costs. Compare the complete process with obtaining and preparing suitable real examples.
12 | DIVERSITY IS MORE THAN IMAGE COUNT
A thousand images can still contain little useful variation if they show nearly identical subjects, backgrounds and viewpoints. Diversity concerns the meaningful range of examples.
A model can inherit gaps and imbalances from its training data. Generating more samples does not automatically fill those gaps or create balanced representation.
Inspect variation relevant to the downstream task. Repeated near-duplicates may add volume without adding much useful information.
13 | THE GAP BETWEEN SYNTHETIC AND REAL
A downstream model may perform well on artificial images yet struggle with real photographs. Differences in textures, backgrounds, lighting or artefacts can create a domain gap .
The learner may also exploit a shortcut. For example, if every generated animal image has the same background pattern, a classifier could learn that pattern instead of useful animal features.
A synthetic image looking convincing to a person does not prove that it supplies the right evidence for a learning task. Evaluate downstream behaviour on relevant real-world examples.
14 | KEEP EVALUATION INDEPENDENT
If a model is intended for real-world images, reserve suitable real examples for evaluation. Do not judge success only on synthetic images from the same generation process used for training.
Separate training, validation and test purposes. Avoid duplicates and near-duplicates across sets, and investigate whether the generator itself may have been trained on intended evaluation examples.
That last issue can be difficult to establish for an external model. Record uncertainty rather than presenting an evaluation as fully independent when the data provenance is unknown.
15 | READ A SMALL EXPERIMENT
Consider an invented classroom experiment using the same downstream classifier design and a fixed real-image test set of 100 items. A baseline trained on real data gets 78 correct. Adding reviewed synthetic training images gives 84 correct.
Accuracy rises from 78% to 84%, an increase of 6 percentage points . That is evidence of improvement in this trial, not proof of improvement for every task or deployment.
Repeat trials, inspect class-level results and account for review cost. If the team keeps choosing configurations based on this test set, it becomes part of development and a fresh independent final evaluation is needed.
16 | SYNTHETIC DOES NOT MEAN RISK-FREE
Generated images can reproduce biases, contain misleading features or resemble training material. Artificial origin does not automatically guarantee privacy, permission or originality.
If people, client designs or protected material are involved, investigate data sources and appropriate use. Keep records showing how data was generated, reviewed and labelled.
Provenance is information about origin and processing history. It supports investigation and responsible decisions; it does not itself prove that every concern has been resolved.
17 | APPLY IT TO VISIONARY STUDIOS
The booklet identifies synthetic datasets as a use of unconditional generation. The studio could investigate this alongside open creative exploration, while recognising that a useful dataset needs preparation and testing.
A downstream classifier is our teaching extension, not a project the booklet states Visionary Studios has already undertaken.
18 | DESIGN A SYNTHETIC-DATA TRIAL
In your book, choose a simple image recognition task. State which useful variations are needed, how synthetic samples would be generated and how labels and image quality would be checked.
Propose a baseline and an independent evaluation. Identify one possible bias, one domain-gap concern and the effort needed to review samples.
Reveal a possible response
For recognising animal categories, compare a fixed real-data baseline with the same learner trained using additional reviewed synthetic examples. Check labels, variation, distortions and near-duplicates. Evaluate on a separate, relevant real-image set and examine each category as well as overall accuracy.
Keep generation records and check whether provenance is sufficient to assess overlap with evaluation data. An improvement in one trial needs confirmation and does not remove bias or permission concerns.
Use the unconditional generation section of the official 2027 booklet alongside these independent teaching examples.
CHECK YOUR UNDERSTANDING | GENERATION METHOD OR DATA USE?