Stability AI Acquires Init ML, Makers of Clipdrop Application | Hacker News Reader

Stability AI Acquires Init ML, Makers of Clipdrop Application

stability.ai

86 pointsdavidbarker20 comments

Man they’re spending so much money. It’s theirs to spend and I wish them the best but I worry about their runway when they’re the only ones really releasing high quality open source models. (I really don’t want to go back to the pre 2021 ML era.)

AI seems to be the only thing you can still easily fundraise for these days, so I don't think that's too much of an issue.

The consesus seems to be that there are too many investors and not enough AI startups to invest in (in SF at least).

Saw a post the other day for an AI meetup where they said they already had too many investors on the waitlist and were no longer accepting more.

Honestly, is there much money to be made from these things long term?

Like I know they will get “better”, but I still struggle generate something that isn’t kind of “weird”, today a horse with a harness melted into its face etc, and even if it generated picture perfect images, there is already a lot of images and art available for cheap, and a lot of competition will come along, but the thing with image generation is there’s nothing really special about it once’s it’s working to a decent enough level ? Will we just have ten services that generate similar stuff ?

While there are lots of obstacles yes I suspect there will be quite a lot of money long term. From my perspective we're in the very early innings of another electricity like revolution. With images you get things like Hollywood being replaced[0] but the much bigger gains come from the seemingly endless list of tasks that these models can approach solving. I feel like I see a new groundbreaking application several times per week at this point[1].

With all that said I'm sure lots of startups will blow up and make no money too. While the overall wave seems very valuable to me, it's hard to predict what will be the Google and what will be the Webvan.

[0] https://www.youtube.com/watch?v=GVT3WUa-48Y

[1] https://twitter.com/DannyDriess/status/1632904675124035585

Can you make a list here of all the groundbreaking applications you've seen?

I’ll be interested to see if Hollywood gets replaced. People actually do enjoy actors, celebrities etc. I’m not sure it will be replaced it will just change.

Also if you think about it, as I said long term, the cost of entertainment will come down so much, because now you just click buttons to generate art, not sure it actually will be worth billions ?

People will still use these things as tools to create more incredible stuff, while the generic kind of boring stuff with a terrible story line will be just absolutely everywhere.

Lastly, I would say that it’s never been easier to make movies and cartoons as it is today but it’s been a pretty long time since I’ve seen a good movie. Movies seem to be about special effects now days.

so if everyone can make great special effects for cheap, maybe they'll have to start focusing on the story again

These art tools are just the first step. Ability for computer to show things live and communicate visually with humans might open up some novel use cases.

"there’s nothing really special about it once’s it’s working to a decent enough level ? "

Art has a pyramid revenue distribution. The top tier of works earn way way more than the 'average' work, this is true for movies, games, TV, novels, illustrations. This is not because of marketing effects, but because top art genuinely elicit far more enjoyment than average art, and audience standards get constantly raised.

After LOTR, Harry Potter, Hollywood tried to create big fantasy movies in the 2000s. But top quality source materials are rare, and with only average scripts, nobody wanted to watch those movies, everything was a flop.

Therefore, there's no diminishing returns in quality. Every extra bit of quality in an AI art engine has a substantial effect.

SD is currently dominant in AI art, because its open source allowed for thousands of passionate tinkerers to finetune the model. There's now like 15 mainstream SD variants, each suited for particular artstyles, there's also LORA databases (Think plugins) that allow people to reproduce every celebrity/famous fictional character in almost perfect quality. Its like having thousands of free developers working for your company, so its going to be extremely hard to compete against SD.

Now SD's biggest problem is its inability to make money off its open sourced success. But its at least cornered the AI art model market. The size of this market is a substantial chunk of the entire entertainment industry, so its easily worth a few billion.

Speaking of runway, I think keeping track of what Runway.ml is doing might be a better move, as they originally made the Stable Diffusion model with the help of a university and which Stability AI then funded for much more computation to make it more powerful. Now they're working on text to video.

Up until today I thought they were a non-profit akin to LAION, funded by some big endowment. Seriously wonder what their path to revenue is.

Can someone shed some light on the ownership structure of Stable Diffusion?

AFAIK the company getting all the attention and funding (Stability.ai incorporated in the UK) is not the one developing the tech. Development appears to originate at the university in Munich / Heidelberg (see e.g. ommer-lab.com).

They’re employing some of the main authors from that university.

stable diffusion 1.0 was a one off collaboration between runway, compvis, and stability ai. (https://research.runwayml.com/the-research-origins-of-stable...) it seems since then the groups have gone their separate ways, with stability releasing an independent stable diffusion 2.0 and runway quite publicly distancing themselves from it https://www.forbes.com/sites/kenrickcai/2022/12/05/runway-ml...

stability is now better funded than runway and about to close a round valuing it at 4bn according to recent reports. there's a central stability team and then it tends to fund/work with other organizations very closely like Laion, CarperAI, Harmonai, and others listed on its website (https://stability.ai/). i am not an insider i just have read a bunch of stuff and this is my best guess right now

Interesting. Emad was just retweeting how Clipdrop’s monocular depth estimator would work well with ControlNet.

I wonder if they’ll open it up or if they’ll keep it proprietary.

Link to related tweet Emad retweeted:

https://twitter.com/sudu_cb/status/1632674745031483392

Interesting, so can this be used to generate a 3d model from a photo?

My understanding is it is a depth map, which is a 2D layer, not a fully 3D map; think of it as a 3D map from a single point of view rendered in 2D using a predefined gradient encoding.

Not one with 360 view if that's what you're thinking. See it as one of those hard halloween masks, from the front they look like the shape of a person's face, but as soon as you look at it from the side/back you just see a thin piece of plastic that has a hollowed out shape.

They're useful as guidance for other image generations to also keep the same structure, or use it to generate a normal map and have dynamic lighting.

Since the terms of the deal are not public, anyone have any idea how to estimate value of the deal. For example: how much Init ML raised prior to its acquisition and common comparable multiples; actually monthly active users and comparable price per user; etc.

Reply on news.ycombinator.com