diffusion model
A class of generative AI models that create data (e.g. images, video) by progressively denoising random noise; often criticized as not 'understanding.'
As defined in Demis Hassabis #2
Other definitions
A generative model trained to gradually remove noise from data or latent representations, capable of learning to generate new samples by reversing the noise addition process.
Yann LeCun: Special Lecture on AI and World Models
Where it comes up · 8
- Jeremiah Lowin: Explaining the New AI Paradigm - [Invest Like the Best, EP.307]4 claims
- Gaurav Misra & Dwight Churchill - Building Captions - [Invest Like the Best, EP.405]2 claims
- A radical theory of consciousness | AI researcher Joscha Bach2 claims
- ChatGPT: The Dawn of Artificial Super-Intelligence | Brian Roemmele | EP 3572 claims
- Gwern — Anonymous writer who predicted AI trajectory on $12K/year salary1 claim
- Demis Hassabis #21 claim
- The Mastermind Behind GPT-4 and the Future of AI | Ilya Sutskever1 claim
- Yann LeCun: Special Lecture on AI and World Models1 claim
Most important claims · 10 of 14Search all claims for “diffusion model” →
- Diffusion models generate video by starting from pure noise and, at each denoising step conditioned on text, progressively revealing structure such as a man in a blue shirt, unlike next-token prediction, which predicts the next word from prior context.Gaurav Misra & Dwight Churchill - Building Captions - [Invest Like the Best, EP.405]
- Stable Diffusion beat DALL-E because it runs diffusion in latent space rather than pixel space; this frees the model from the spatial constraints of the image during generation, so it can merge image content first and then settle details like five fingers at the end, which gave it much greater expressive power.Jeremiah Lowin: Explaining the New AI Paradigm - [Invest Like the Best, EP.307]
- Modern diffusion latent spaces are too large for a single slider per attribute, but this does not make the models less interpretable; it only makes them more complicated, and a machine can find the combination of numbers that yields an attribute such as a bigger smile.Jeremiah Lowin: Explaining the New AI Paradigm - [Invest Like the Best, EP.307]
- GAN and diffusion models think differently because of their loss functions: GANs face an adversarial loss giving them an incentive to hide difficult things, so they are 'scared' and put hands off-screen and can't think about them, whereas diffusion models have no such incentive and do attempt hands, producing monstrous Cthulhu-esque distortions instead.Gwern — Anonymous writer who predicted AI trajectory on $12K/year salary
- Joint embedding architectures that do not attempt to reconstruct perform much better than reconstruction-based methods (autoencoders, VAEs, masked autoencoders, diffusion models) for learning image representations that transfer well to downstream tasks.Yann LeCun: Special Lecture on AI and World Models
- Veo 3 demonstrates a form of understanding—not deep philosophical understanding but intuitive physics akin to a human child's—by accurately predicting coherent next video frames including liquids, materials, and specular lighting, reverse-engineered purely from watching videos, which challenges the theory that physical understanding requires embodied interaction with the world.Demis Hassabis #2
- Text models are in the roughly 400-billion-parameter range, while diffusion video models are in the tens of billions of parameters, and Meta's MovieGen was about 30 billion, so video has far more room to scale.Gaurav Misra & Dwight Churchill - Building Captions - [Invest Like the Best, EP.405]
- Stable Diffusion's visual model contains more visual knowledge (all celebrities, dinosaurs, spaceships, artistic styles) at higher fidelity than a human brain, yet fits in 2 gigabytes, suggesting that roughly 80% of what the human brain does is compressible into this volume.A radical theory of consciousness | AI researcher Joscha Bach
- You can use ChatGPT to generate sophisticated prompts for diffusion models like Midjourney to create complex artwork, turning one AI (language) into a tool to instruct another AI (visual generation), with the human as a creative director.ChatGPT: The Dawn of Artificial Super-Intelligence | Brian Roemmele | EP 357
- Diffusion models are distinct from large language models and can be applied to visual generation, enabling AI systems to create images that no human has ever seen by synthesizing novel combinations.ChatGPT: The Dawn of Artificial Super-Intelligence | Brian Roemmele | EP 357