Stable Diffusion Chaos — Stability AI Unscrambling Eggs
So I forgot one saying when I reflected on AI art’s sudden impact. You can’t unring a bell and you can’t unscramble an egg.
This past week has been interesting watching Stability AI — the ones who paid for the training of the Stable Diffusion model — bounce around off the hallway walls like a belligerent drunk.
Pop my popcorn and put it in an extra-large tub, please, this is looking like one fun mess in the making. And, yes, everyone saw this coming.
First of all, Reddit. There is a Stable Diffusion Reddit that Stability AI came in and plowed through the china in the shop. Later a new Stability AI hire said, “Hey, our bad, we’re cool.”
Then on 10/20/2022 out of nowhere the long desired where-is-it-you-said-it-was-going-to-be-released-soon V1.5 of the Stable Diffusion model dropped on Hugging Face. Then it showed a bright red warning that this model was part of a legal takedown request. From Stability AI. Then Stability AI said, “Hey, our bad, we’re cool.”
Then on 10/21/2022, Daniel Jeffries from Stability AI wrote: Why the Future of Open Source AI is So Much Bigger Than Stable Diffusion 1.5 and Why It Matters to You (substack.com) .
Oh, we are definitely not cool.
So on one side you have open research including Runway Research with distribution on Hugging Face and the other side you have Stability AI. The creators + innovators side and the funder side. And Stability AI is looking at a bajillion valuation to protect.
For what? A mess. An open source, open internet mess of scrambled eggs, spilt milk, and rung bells. And popped out corks. Threats of regulations, lawsuits, and more. Oh my.
One mess is the fantastic Automatic1111 local distribution for running Stable Diffusion at home. Around the same time Stability AI took over moderation of the Stable Diffusion Reddit, Automatic1111 was kicked to the curb from Stability AI’s ecosystem. We don’t know why (stolen code that wasn’t stolen), but probably the fact that Automatic1111 is tightly coupled with erotic Waifu model generation community doesn’t help.
Open means open and people can add new training to the Stable Diffusion world or use its existing training for Playboy photo shoots. Of anyone. Look, I tagged some of my postings about Stable Diffusion on Twitter with #StableDiffusion. Then one day I innocently went to that tag and goodness if I was wearing pearls I would have clutched them hard. It’s full of young ladies with impossibly top heavy proportions — having a friend in high school whose vertebrae were fusing due to her burden all I see is pain.
All Stability AI seemingly sees is a blocker to raising capital and a grounded reputation.
Let alone all the darker mess that people are training extensions to the model to do. I don’t want even to think about that.
This apparently reactive public display of randomness by Stability AI to secure funding and reputation reminds me of management policies I ran into as a development manager. I struggled against policies and attitudes that stemmed from “we have to do this or otherwise the low performers will take advantage of the company.” Leadership and HR would hesitate on great benefits because they had visions of precedence allowing low performers to dance around enjoying long periods for bereavement leaves or such. No, you manage to your strengths to support and retain your top performers, not degrade everyone’s experience because of fears you have of people you should move on anyway.
How does this apply? Stability AI should own the chaos. It should own the disruption. It should ride the lightning. Not manage to the low expectations of potential funder’s meekness in some sort of cash haul — such is the path of mediocracy. Have a strong, coherent vision and take funding from those who unabashedly support it. Be bold with disruption and be the icebreaker of AI/society friction.
Baring that: friggin’ be consistent. Maybe Mr. Jeffries’ missive is the start of that, I don’t know. Next week some other new Stability AI hire might come up with a different perspective and say, “we’re cool.” As I chomp my popcorn I can only hope that Stability AI’s vision (oy, I tried to watch that recent video but it was too painful) is in fact visionary and not just a scramble to hustle while the hustling is good.
The openness of the AI model is its disruptive strength right now. Let’s be honest: it’s not due to its artistic merit in comparison to what MidJourney can produce. And if Stability AI attempts a lobotomy of sorts for the model (does that include addressing Greg Rutkowski’s pain?) I can only expect that a distributed community training system to leverage all these gaming GPUs will take hold. I expect it’s being written as I type.
Time over many tubs of popcorn will tell. 🍿
