Generative Models: Plausible Does Not Mean Observed
A generated image can look convincing without depicting an event that ever happened.
Generative Models: Plausible Does Not Mean Observed explained visually
A generated image can look convincing without depicting an event that ever happened. English captions are optional and start off. The complete transcript follows the written lesson.
What you will learn
A generated image can look convincing without depicting an event that ever happened.
Learn a distribution
A generated image can look convincing without depicting an event that ever happened. Generative models learn patterns that support producing new samples. That capability is useful for creative work and some data workflows, but it is different from measuring a scene or recovering a uniquely correct missing detail.
Compare two generation ideas
A generative adversarial network trains a generator alongside a discriminator in an adversarial objective. Diffusion models learn a denoising process associated with progressively corrupted data and use a sampling procedure to generate outputs. These descriptions hide many architectural choices, but the central point is that generation depends on learned distributions and conditioning, not direct observation of the produced scene.
Preserve the distinction from evidence
For restoration, inpainting, or synthetic training data, evaluate fidelity and usefulness separately from appearance. Generated objects may have incorrect structure, and synthetic datasets may omit important real-world variation. Keep real evaluation data independent, label synthetic outputs, and preserve original images when they are evidence. A visually pleasing result is a reason to inspect further, not a sufficient accuracy claim. Evaluate diversity as well as visual quality: convincing samples can still omit important variation.
Worked example
If a model fills an occluded vehicle region with a plausible wheel, that wheel may differ from the actual hidden wheel. The output can be useful as an illustration or hypothesis, but it should not be reported as a verified observation of the concealed object.
Quick check: test your understanding
Try it yourself: How would you test whether synthetic images improve a detector?
Check the reasoning
Compare training with and without the synthetic data while keeping a representative real held-out test set fixed and independent.
Does realism prove that a generated scene is factually accurate?
No. Realism concerns appearance; factual accuracy requires evidence about the actual scene.
Further reading
Dive into Deep Learning — Aston Zhang, Zachary C. Lipton, Mu Li, and Alexander J. Smola.
Deep Learning — Ian Goodfellow, Yoshua Bengio, and Aaron Courville.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 79 transcript233 words · Video transcript · Matches the published narration and captionsOpen transcript
A picture can look real and still show something that never happened. That distinction becomes essential when we work with generative models.
These models learn patterns that help them produce new samples. Instead of only assigning a label to an image, they can generate an image or complete part of one.
Different approaches learn this behavior differently. A generative adversarial network, often called a G A N, trains a generator together with a discriminator. The generator tries to produce convincing samples, while the discriminator learns to distinguish generated samples from real training examples.
Diffusion models use a different idea. Training involves progressively corrupted data and learning to remove noise. A sampling process then uses learned denoising steps to produce an output.
Neither approach turns a plausible result into a verified observation. If a model fills a hidden part of a car with a convincing wheel, it has not discovered exactly what the real hidden wheel looked like.
Evaluation therefore needs more than attractive examples. Check quality, diversity, and usefulness for the intended task. A generator that produces one impressive kind of image may still miss important variation.
If generated images are used for training, keep real evaluation examples independent. Label synthetic outputs and preserve original evidence. Generation can be useful for illustration, design, and carefully tested data workflows. Just keep a clear distinction between what the model created and what was actually observed.