Atlas record: https://www.aitimeline.dev/en/events/evt-0010/
Research · March 12, 2015
A generative process of noising and denoising was described.
It offered a probabilistic route to generating data through repeated denoising.
This early paper is an antecedent, not a claim about every later image generator.
arXiv · Published March 12, 2015
Proceedings of Machine Learning Research (ICML 2015, 32nd International Conference on Machine Learning) · Published 2015
Both describe image generation learned from data; the training mechanisms differ.
The same move: a generative model with an explicit probabilistic justification.
The same forward and reverse diffusion described in 2015; this paper builds training on a weighted variational bound on top of it.
Both records concern generative image models; that paper is earlier work in the field, not the basis of this particular model.
The model belongs to the diffusion family.
Video is generated by the same process of gradually removing noise.
The structure is generated by a diffusion module in place of the earlier geometric block.
Generating an image frame by frame, but each next one depends on the viewer's action.
Both outstanding papers work inside this family: one on order in diffusion language models, the other on the accuracy of sampling from diffusion models.