Stable Diffusion used the same architecture as latent diffusion but was trained with more compute

Overview Debates Related claims Topics

Stable Diffusion models tend to recreate recognizable facial features from training data rather than generating wholly novel facial structures from scratch

The diffusion component of a robotics model's architecture refines its output to produce smooth, precise, and fine-grained motions

A diffusion model iteratively transforms random noise into a high-dimensional, well-defined output

The Dyna2 world action model uses a diffusion transformer architecture

Diffusion architectures are effective for visual domains like generating images, videos, and robot actions

The latent diffusion paper was co-authored by Patrick Esser while he was a researcher at Runway

Latent diffusion was a research collaboration between Runway and the Conviz lab

Stability AI financed the large-scale training run that produced the Stable Diffusion model

Runway released Stable Diffusion 1.5 to maintain the project's open-source research origins

Gen-3's architecture uses diffusion transformers instead of the convolutional nets used in earlier models

Atlas is an autoregression diffusion model designed for next-frame prediction.

Diffusion models operate by learning to reverse a process that destroys structure and increases entropy.

The first academic paper on diffusion models referenced non-equilibrium thermodynamics in its title.

The physics concept of counter-diabatic driving can be applied to improve the generation process in diffusion models.

Transformers and diffusion models are the two primary architectures in modern AI.

Chai 2 is described as an all-atom diffusion model.

The Chai 1 model's architecture consists of a tokenizer, a transformer, and a diffusion model that outputs a 3D structure.