A brief history of diffusion models

Think of dropping a sugar cube into hot coffee. As it dissolves, the sugar particles spread out randomly until the sweet cube is completely lost to entropy. In the physical world, there is no way to put that cube back together.

Computer science figured out how to reverse the physics. By slowly adding random static to an image until it becomes pure noise, and then training an AI to step-by-step remove that noise, researchers created diffusion. Over the past year, this exact mechanism quietly took over creative tools at Canva and DeviantArt through systems like OpenAI's DALL-E 2 and Stable Diffusion.

Yet image generation is just scratching the surface.

Leaving brittle GANs behind

Before diffusion ruled the roost, generative adversarial networks (GANs) were the undisputed kings of synthetic media. They powered the early wave of deepfakes and Nvidia's StyleGAN, which famously generated photorealistic faces out of thin air.

GANs pitted two networks against each other: a generator crafting fake samples and a discriminator trying to catch them. In theory, both got smarter together. In practice, the training was notoriously fragile. Generators often collapsed into outputting the exact same blurry mess, and scaling them required monstrous amounts of compute.

Diffusion changed the math. While the core idea kicked around for nearly a decade, it only became truly practical when OpenAI introduced CLIP. Think of raw noise as a block of uncarved marble: CLIP acts as the master sculptor looking over the novice's shoulder, scoring each denoising step against your text prompt until an image emerges.

Beyond art: from beats to biology

Researchers quickly realized that any data representable as structured noise can be synthesized using diffusion.

In audio, Harmonai trained diffusion models on hundreds of hours of songs, while a viral hobby project called Riffusion converted audio into visual spectrograms to generate tunes with standard image diffusion models.

The real frontier, however, is computational biology. Generate Biomedicines is using diffusion to assemble novel proteins by unraveling amino acid chains and stitching them back together. A University of Washington team trained a diffusion model to design proteins from scratch, successfully producing one that binds to parathyroid hormone better than existing commercial drugs.

Meanwhile, the open-source initiative OpenBioML is developing DNA-Diffusion. The goal is to generate cell-type-specific regulatory DNA sequences simply by typing prompts like targeting gene activation in heart and liver tissues without touching the brain.

Why it matters

For creators and indie builders, diffusion has flattened the cost of generative experimentation. Projects like Riffusion show that you do not need a multimillion-dollar lab to create novel applications; clever formatting of inputs is often enough. The catch right now is legal and technical: as researchers have pointed out, these models occasionally regurgitate their training data, leaving commercial adopters exposed to IP disputes.

For the broader tech landscape, diffusion is shifting the frontier from consumer novelties to hard science. Proven success in drug discovery and synthetic biology shows that generative architectures can solve high-stakes physical problems rather than just mimicking digital art styles.

We think diffusion's real triumph was ditching the adversarial instability of GANs for a predictable thermodynamic step-by-step process. It remains compute-heavy and will likely face competition from even leaner architectures in the years ahead, but right now, it is undeniably the most versatile engine driving generative AI.