technologybriefs
9:18in productionCh. 1 · What it actually does/ 9:18 · ceiling 15 min
Models · AI

Diffusion model

Diffusion models don’t generate images—they reverse entropy, one noisy step at a time.

Diffusion models are a 2015 method for sampling from complex distributions using thermodynamic-inspired noise reversal. They work by learning to undo a diffusion process—turning noise into data, step by step. They succeed in computer vision tasks as of 2024. They do not generalise beyond vision. They require iterative sampling, trading speed for quality. Their mechanism is physical in inspiration, statistical in execution, and narrow in current utility.

ModelsAIArtificial intelligenceMachine learningAlgorithmsArtificial intelligence across the network →
Chapters & takeaways4
  1. 0:57
    What it actually does

    They are not magic: they learn how to undo noise, not how to imagine.

  2. 2:12
    How it walks backward

    Each new sample is a physical-like trajectory—not a lookup or a guess.

  3. 3:39
    The 2020 upgrade

    DDPM improved training, but sampling remains a spectrum of speed versus fidelity.

  4. 5:10
    Where it sticks

    As of 2024, they are vision-only tools—not language, audio or control systems.

Worth your time?

Yes. Study the whole thing.

3.5/ 5
What works
  • learning diffusion processes for given datasets
  • generating new elements distributed like originals
  • image denoising, inpainting, super-resolution, generation and video generation
What does not
  • generalise beyond computer vision
  • deliver real-time sampling without trade-offs
  • replace other generative paradigms outside image/video domains
Study it if
  • computer vision engineers
  • researchers building image synthesis pipelines
  • teams deploying high-fidelity generative tools where latency is secondary
Skip it if
  • real-time interactive applications
  • low-power edge devices
  • non-vision modalities without adaptation
The written brief1 min read

What it is and the problem it solves

Diffusion models are a method introduced in 2015 to train a model that can sample from a highly complex probability distribution. They solve the problem of generating realistic data points when the underlying distribution is intractable to describe directly.

How it works

Diffusion models model data as generated by a diffusion process: a new datum performs a random walk with drift through the space of all possible data. They use techniques from non-equilibrium thermodynamics, especially diffusion. The 2020 DDPM variant improves training via variational inference.

What works

They reliably generate new elements distributed similarly to the original dataset—by learning a diffusion process for a given dataset. As of 2024, this works robustly for image and video generation, denoising, inpainting and super-resolution.

What does not

They do not solve sampling efficiency at scale. A trained diffusion model can be sampled in many ways, with different efficiency and quality—meaning speed, compute cost and fidelity trade off directly. No claim states they are fast, lightweight, or broadly deployable beyond vision.

What it changes

They shift generative modelling from autoregressive or GAN-based paradigms to iterative reverse-diffusion sampling. This establishes a new standard for distribution matching in vision—but only where high-fidelity synthesis matters more than latency or energy cost.

Is it worth your time

Yes—if your work involves computer vision tasks like image generation, denoising, inpainting, super-resolution or video generation. As of 2024, that is their dominant application. They are not general-purpose generative tools outside those domains.

Same field · Models4 of 5
8:14
Gaussian splattingGaussian splatting is a volume rendering technique introduced by Lee Westover in the early 1990s and revitalised in 2023 by Inria for real-time radiance field rendering. It uses differentiable, anisotropic, explicit 3D Gaussians integrated from sparse camera-calibrated points. It is rasterisation-based, GPU-optimised, and avoids surface or line primitives. It enables fast photorealistic reconstruction from few images—but offers no mesh output, no topological guarantees, and limited editability.
9:11
Generative adversarial network2014GANs are an adversarial framework for unsupervised generation. They work by pitting two neural networks against each other: a generator learns to produce data that fools a discriminator, which itself improves at spotting fakes. This indirect training yields samples statistically aligned with the training set—including photorealistic images that pass superficial human inspection. But GANs offer no guarantees on semantic coherence, stability, or interpretability. Their value lies in enabling distribution learning without labels—not in reliability, control, or fidelity.
9:28
LoRA (machine learning)LoRA makes fine-tuning large language models radically cheaper to train without slowing inference—by replacing full-weight updates with low-rank matrix additions. It works. It scales. It merges. But it assumes weight updates are low-rank—and that assumption fails where tasks demand dense, global parameter shifts.
9:29
Reinforcement learning from human feedbackRLHF replaces hand-coded rewards with learned ones—using human rankings as ground truth. It enables alignment at scale for language models, but only within the narrow bounds of the annotation task. It is widely used, not universally reliable.
Up next in Technology

Direct air capture

· 9:28

DAC removes CO₂ from thin air—but not at scale, not cheaply, and not yet.

Play nextBack to Models
9:28