Thread by @IntuitMachine

It is extremely instructive to follow the exponential growth of diffusion models in AI art. At present 3 major players @openAI with Dall-E, @midjourney founded by @DavidSHolz, and @StabilityAI founded by @EMostaque. Each player has taken a different path in their release.

More capable systems exists from Google (i.e. Imagen imagen.research.google ) and Facebook (Make-A-Scene ai.facebook.com/blog/greater-c…), but these aren't available to the public.

I want to discuss how players release their AI products into the wild. #dalle2 has been available since July 20th after opening up a restrictive waitlist 2 months earlier. I had the chance to refine my prompting skills since then:

https://twitter.com/IntuitMachine/status/1540796489579896833

When you first access a tool like this, you don't know how to prompt what you imagine. You learn either through experimentation or through examples of other users. Twitter was the medium where you could learn from others. @hardmaru generated a ton of interesting ideas.

The other tool that was also out there was @midjourney. I've not had access (I'm not an artist) until recently. I was surprised to discover that you can only use it via Discord. I find Discord to be too noisy and always thought using Discord was for convenience and not design.

@StabilityAI mirrored the same roll-out on Discord. They employed a superior algorithm based on Perceiver-IO that serendipitously could run on a single GPU.

https://twitter.com/IntuitMachine/status/1554128453158461442

Stable Diffusion is trained on an aesthetic subset of images (i.e. aesthetics score of 6 or more). Thus unlike Dall-E 2, it's biased towards a certain kind of aesthetics. huggingface.co/datasets/Chris…

We often focus too much on how AI is trained but overlook the design of its community of users. There is a clear difference in how people engage with these tooling, and I would like to discuss these differences.

Stable Diffusion came out first on Discord. They've shut it down once they open sourced the neural network. The interesting feature of SD was that you could borrow the seeds used by other users and recreate the images or variants of the images. You can't do this with Dall-E 2.

Thus the Discord server offered a means for users to reuse images created by other users. With a seed, you can generate new variants by tweaking the original prompt. Discord's shared space thus became a hive-mind brainstorming space.

This reuse of images is also available with Midjourney, but the engagement is different. Instead a user can only upscale or create variants of the generated image. The seeds are not made available. The consequence of this is you can always discover an image's ancestry.

Ancestry is very important because it's necessary to recreate images. Recreating images is next to impossible in Dall-E 2.

https://twitter.com/IntuitMachine/status/1550170265321095171

David Holz @DavidSHolz Midjourney takes the genetic approach of ArtBreeder and fuses it with a Stable Fusion network. It's an open-ended ecosystem for artists to discover novel ways of expression. artbreeder.com/beta/image/93c…

In contrast, Stable Diffusion open source leads to another kind of open-ended evolution. Programmers are now tweaking the algorithm to add new capabilities such as img2img, inpainting, reverse text, etc. This can only lead to the development of more innovative workflows.

It feels like drinking from a firehouse when you enter the Discord servers of Midjourney or StabilityAI. It's too much to grasp too soon! Many are in panic and there's a noticeable interest in "AI alignment." David Sholz says its like swimming in water. theverge.com/2022/8/2/23287…

You can either drown by consuming too much water or go with the flow and accept the richness of collaboration. In AI, we must accept that new technologies lead to new ways of design. Therefore how we design human interaction enabled by AI should not be overlooked.

The utter failure of the AI and AI Alignment community is the lack of realization of the importance of collaborative design. Humans, after all, have designed our societies just as we've designed our technologies. Treating these as separate things solves just half the problem.

So here we are today, inundated with AI art, yet we are still grasping in the dark as to how to deal with it. We can't deal with it unless we change the framing of what it means to be human. An area of study that civilization has been ignoring for too long.

@threadreaderapp unroll