First visual model trained on fully open data coming. Like for text, the striking thing is not only that it is possible in the first place but that it brings so much more variety and diversity of outputs. There’s an entire world beyond web archives.

Taking shape.

More Public Diffusion samples! It’s still very early on (<33% done), but I’m optimistic with the quality of these outputs. Some interesting examples in 🧵, all prompts in the alt text. My favorites:

The model also seems to be good at differentiating between mediums. These images used the same prompt but changed the medium. One in watercolor, and the other in charcoal:

The king of YouTube AI image tutorials covers Public Diffusion ❤️ Thanks Olivio. To confirm the weights will be unconditionally public.

Thank you, Olivio Sarikas for this excellent introduction to the Public Diffusion project!

✨New Report✨ Our data ecosystem audit across text, speech, and video (✏️,📢,📽️) finds: 📈 Rising reliance on web, synthetic, and YouTube data. 🛑 80%+ datasets carry hidden restrictions. 🌍 Relative representation in languages and creators has not improved for 10+ yrs.

I sat down with CEO Jordan Meyer and CTO Nick Padgett to talk all things Public Diffusion, from fine-tuning, to prompting like a photographer and what’s next for beta testers. Link to the interview, and some choice quotes, in 🧵

Interview with Jordan Meyer and Nick Padgett who are building Public Diffusion, with some more beautiful examples from the model in training The colorspace and contrast is stunning

Happy AI breakthrough day! 🎄 I'm so proud this beautiful book is finally out! This was a huge group effort and covers much of Mat Dryhurst and my thinking around Machine Learning and Art Thanks to our editors Eva Jäger Caroline Busta, designer Eric Hu and to the inimitable

It’s surreal that I be the subject of controversy over the provenance of data used in my art (about the provenance of data) Your talking points on this issue are my throwaway observations from 2018 Joke world

Im also training a diffusion model on the PD12M (public domain) dataset Generated in 2 seconds on my 3090 btw (768x1536) and the code is unoptimized

If only someone had built this two years ago

How to build AI leadership in the U.S? It is not about the chips. It is about the people! I spoke about AI interpretability at ntird.gov last week. (NTIRD is the joint program between 23 federal agencies that coordinates government tech investments.)

OpenAI announces they will sign the EU code of practice nb UK policymakers, this includes opt-outs There needs to be one universal standard that aids both sides by omitting duplicates - as duplicates ingested by models will be a legal pain Happy to contribute to that