Generative A-Eye
This newsletter will be back shortly, and will be much more frequent - daily, in fact, and concentrating on a digest of the hottest new research in human image synthesis
For the last two and a half years I’ve been writing features and pounding the science beat, providing content for the amazing AI VFX company Metaphysic.ai.
The resultant 150+ articles, which are mostly write-ups of some of the most potentially interesting new science papers related to human image synthesis, can currently be seen at https://blog.metaphysic.ai/ (however the page is a little broken, and doesn’t show the earliest posts, including the bigger features, and will likely be changed soon - but a complete list can be found at my portfolio site at https://martinanderson.ai/#METAPHYSIC).
I’ve seen things you people wouldn’t believe.
Roy Batty
My domain knowledge of the state of the art in AI-based image synthesis for VFX has been significantly expanded by participation in production dailies, and insights provided by the genius teams at the company; dedicated people tackling what we all know are some mighty obstacles in achieving human synthesis through machine learning methods, rather than traditional CGI approaches.
Though I can’t discuss anything specific from that time, I do now have a much clearer overview of the challenges ahead – and the extent to which this year’s headline-grabbing T2V systems (Sora, Kaiber, Luma, Kling, Gen3, et al.) haven’t yet succeeded in solving these roadblocks.
I’ll be returning to these topics once again at unite.ai, shortly. Which brings me to the topic of this post.
AI Surveillance
Over the last four years, at Unite and then Metaphysic, I have kept a constant eye on the image synthesis and computer vision research scene. It’s quite a task, since Arxiv’s Computer Vision section alone can generate 2000+ submissions a week, in high conference season.
Therefore I would take note of those new entries that seemed most promising and compile them in a succession of voluminous Google Docs - enormous collections of select proposals for possible articles (since there wasn’t possibly time or resources to write them all up).
These entries comprised titles, summary selections, links, an exemplary illustration, and my brief notes on their potential significance.
Since I will be continuing this practice for Unite; since I will remain unable to fully write up all the papers that catch my attention; since these compilations often proved useful to time-starved researchers at Metaphysic; and since I have to deposit the candidates somewhere1 – I have decided to revive this long-dormant newsletter as a daily ‘best of’ update from the research scene, under the moniker Generative A-Eye.
If you don’t want to receive a newsletter that frequently, please unsubscribe2.
This doesn’t mean that this newsletter will not occasionally have some under-the-fold content from me that doesn’t fit in elsewhere; but it will now appear beneath the day’s research selections (on the occasions where I have anything extra to say, outside of my other outlets).
Martin Anderson
https://martinanderson.ai
And am not the biggest fan of Twitter, though I have a neglected account there.
I haven’t explored SubStack in a few years, but I don’t think they allow subscribers to receive a weekly digest of a daily newsletter.