Every model discussed on this blog so far has shared a common underlying job: looking at something and producing a judgment about it. A CNN looks at an image and decides what it contains. A classifier looks at patient data and predicts a diagnosis. An evaluation metric looks at a model’s output and scores how good it was. This entire family of models is sometimes called discriminative, because their fundamental task is discrimination in the statistical sense, drawing boundaries between categories or predicting an outcome from an input. Generative models flip this entire premise around, and understanding that flip is the key to understanding one of the most consequential branches of modern artificial intelligence.
What Makes a Model «Generative» Rather Than «Discriminative»
Generative models are a class of machine learning models that learn the underlying distribution of a dataset in order to generate entirely new data samples that resemble the original input data, in contrast to discriminative models, which instead predict labels or outcomes given a specific input. Rather than answering the question «what is this,» a generative model is built to answer a much stranger and more ambitious question: «if I understand deeply enough what data like this tends to look like, can I produce something new that plausibly could have come from the same source?»
At their core, generative models are essentially trying to answer the question of how to produce something new that still fits the same underlying data distribution, given everything the model has learned about that data. This distinction matters because it reframes the entire learning problem. A discriminative model only needs to learn the boundary between categories. A generative model needs to learn something closer to the full shape of the data itself, which is a considerably harder and more information-rich thing to capture.
This is also where the «probabilistic» part of the topic becomes central rather than incidental. Generative modeling has become a central approach for learning complex probability distributions and producing data that closely resembles existing samples, and this probabilistic grounding has significantly shaped the broader landscape of machine learning research over the past several years. A generative model is not memorizing specific examples from its training data. It is learning an approximation of the probability distribution those examples were drawn from, which is precisely what allows it to sample entirely new points from that same distribution rather than simply reproducing what it has already seen.
Variational Autoencoders: The Gentle Introduction
The story of modern generative modeling usually starts with autoencoders, and specifically with a probabilistic twist on them called the variational autoencoder, or VAE. A standard autoencoder learns to encode input data into a compressed, lower-dimensional representation called a latent space, then decode that compressed representation back into something resembling the original input. A variational autoencoder takes this same basic structure and adds a genuinely important probabilistic refinement. Rather than encoding an input into a single fixed point, a VAE learns to encode it as the statistical parameters of a probability distribution, which leads to a meaningfully more useful and better structured generated result.
This distinction matters more than it might initially sound. VAEs use an encoder to map input data into a latent Gaussian distribution rather than a single fixed point, and a decoder that then reconstructs data by sampling from that distribution. Because nearby points in this learned latent space tend to correspond to gradually and smoothly changing outputs, VAEs produce a smooth latent space where small changes in the underlying latent variables yield correspondingly smooth changes in the generated output, a property especially useful for creative tasks like blending or interpolating between two different generated results.
VAEs come with real practical advantages: they train in a stable, relatively straightforward manner, require no labeled data since the learning process is fundamentally unsupervised, and work well for tasks like anomaly detection, where an unusually large reconstruction error signals that a given input does not fit the pattern the model learned from normal data. Their main drawback, one that motivated much of what came next in this field, is quality. VAEs often produce noticeably blurrier outputs compared to some of the alternative approaches that followed, and can struggle to capture fine-grained detail in genuinely complex data.
Generative Adversarial Networks: Two Models Competing Against Each Other
The next major leap in generative modeling took a completely different, almost adversarial approach to the problem, and it remains one of the more conceptually elegant ideas in modern machine learning. A generative adversarial network, or GAN, comprises two competing neural network modules trained together: a generator, whose job is to create a convincing fake sample, and a discriminator, whose job is to determine whether a given sample is genuinely real or artificially generated.
The two networks are trained simultaneously in direct competition. The generator gets better at producing convincing fakes specifically by trying to fool the discriminator, while the discriminator gets better at spotting fakes specifically by trying to catch the generator’s mistakes. Over the course of training, this adversarial back and forth pushes both networks to improve together, ideally converging toward a generator capable of producing outputs the discriminator can no longer reliably distinguish from the real thing.
Comparing the two major early approaches directly, VAEs tend to exhibit greater training stability, while GANs tend to excel specifically at producing more visually realistic, higher fidelity images. That sharper visual quality comes at a real cost, however. GAN training is well known to be unstable, prone to a specific failure mode called mode collapse, where the generator learns to produce only a narrow, repetitive range of outputs rather than genuinely capturing the full diversity of the original data distribution.
Diffusion Models: The Current State of the Art
The generative modeling landscape shifted again, more recently and more dramatically, with the rise of diffusion models, which now dominate the highest quality image generation systems in active use. Diffusion models generate data through a genuinely different mechanism: they iteratively refine random noise through a reverse diffusion process, gradually transforming pure statistical randomness into a coherent, high-quality output over many small denoising steps.
The underlying process runs in two directions. A forward process gradually adds Gaussian noise to a real training image over a sequence of steps, progressively destroying its structure until essentially nothing recognizable remains, while the reverse process trains a network to undo this destruction step by step, learning to remove a small amount of noise at each stage. Once trained, generating a brand new image involves running this reverse process starting from pure random noise, gradually denoising it across many steps until a coherent image emerges that was never part of the original training data but plausibly could have been.
This approach directly addressed the biggest weaknesses of what came before it. Compared to GANs, which are prone to unstable adversarial training dynamics and issues like mode collapse, diffusion models offer a considerably more stable training procedure, and unlike VAEs, which often struggle to produce genuinely high-quality samples due to their reliance on an approximate probabilistic structure, diffusion models tend to combine stability, output diversity, and high fidelity all at once. This combination of strengths is precisely why they have taken over as the dominant approach. Diffusion models have effectively taken over the generative image landscape, now serving as the foundation for most modern text-to-image and image-to-image systems, while GANs have been relegated to a smaller set of niche applications like image upscaling and style transfer.
The main tradeoff, and it is a significant one, is computational cost. Generating a single image through dozens or even hundreds of sequential denoising steps is inherently more computationally expensive than a GAN’s single forward pass through its generator. Much of the recent engineering effort in this space has gone into addressing exactly that tradeoff, often by running the diffusion process in a smaller, compressed latent space rather than working directly on full resolution pixels, borrowing the same compression idea that made VAEs efficient in the first place.
Where This All Connects Back to Earlier Topics on This Blog
This entire family of models sits at a genuine intersection of several topics already covered here. The underlying architecture inside many modern diffusion models still leans on convolutional structures discussed in the earlier CNN post, applied not to classify an image but to predict what noise should be removed from it at each step. The training process, particularly for GANs, echoes the reinforcement learning post’s discussion of two systems locked in a feedback loop, though the mechanics differ meaningfully, since a GAN’s generator and discriminator are trained through direct gradient-based competition rather than reward signals accumulated over time. And the underlying philosophy, learning a probability distribution well enough to sample genuinely new, plausible points from it, is really just the machine learning post’s core idea, learning patterns from data rather than being told explicit rules, pushed to its most ambitious conclusion: not just recognizing what already exists, but producing something that never did.
Where This Field Is Heading
A defining trend across the most recent developments in this space is the integration of large language models with image generation systems, using the language model’s ability to interpret prompts more precisely to guide and improve the fidelity of what gets generated, a direct connection back to the earlier post on this blog about the fundamentals of LLMs. Generative modeling has also expanded well beyond images alone, reaching into audio synthesis, molecular design for drug discovery, and increasingly sophisticated text generation, all built on the same underlying probabilistic foundation: learn the true shape of the data well enough, and creation becomes just another form of sampling from what the model has understood.
What started as a fairly narrow academic question, how to build a model that learns a probability distribution rather than just a decision boundary, has grown into one of the most visible and culturally significant branches of artificial intelligence in existence today. And underneath every headline-grabbing image or video these systems produce sits the same quiet mathematical idea that has run through this entire family of methods from the beginning: understanding a distribution well enough to draw something genuinely new from it.
By: Max Johnson B.
Deja una respuesta