Seems the CEO/Founder of Runway has replied that there hasn't been any leaks and it's a proper release, made as permitted by the license and therefore should be online soon again.
> Cris here - the CEO and Co-founder of Runway. Since our founding in 2018, we’ve been on a mission to empower anyone to create the impossible. So, we’re excited to share this newest version of Stable Diffusion so that we can continue delivering on our mission.
> This version of Stable Diffusion is a continuation of the original High-Resolution Image Synthesis with Latent Diffusion Models work that we created and published (now more commonly referred to as Stable Diffusion). Stable Diffusion is an AI model developed by Patrick Esser from Runway and Robin Rombach from LMU Munich. The research and code behind Stable Diffusion was open-sourced last year. The model was released under the CreativeML Open RAIL M License.
> We confirm there has been no breach of IP as flagged and we thank Stability AI for the compute donation to retrain the original model.
Edit: According to Discord messages from Emad Mostaque (founder of Stability AI) they have taken back the takedown request and seems this is the official release of the Stable Diffusion model version 1.5.
Sounds to me like runway released it without consulting stability, called it “1.5” - which according to the license they’re allowed to do but pretty scammy since emad had hyped a model with that label. And now stability is deciding to call this the official release to be nice to runway and avoid a general PR thing and community infight.
1.5 was apparently held back by Stability for weeks.
Runway finally decided to just release it.
Stability requested a take down, and here you see the Runway CEO telling Stability in no uncertain terms "this ours to release, we created it, it's under an OS license, you don't hold any IP rights here; all you did was provide compute"
If anything this is a pretty stern rebuke of Stability and a sign of considerable disagreement between the two parties.
Hilarious. So a VC darling $100mil investment company is just a server farm dishing free compute donations around to get some good PR and doesnt own anything.
This is exactly why Stability's role as a middleman in these model releases is an indefensible business position. The research underlying stable diffusion does not belong to them, so any company which wants to fund the training of the next iteration of these models can easily step in and fill their role. And since Stability itself is not pioneering this line of research (rather, they are just facilitating the training and distribution of models), they will always be lagging behind other companies which employ the researchers themselves (i.e. Runway's apparent position in this debacle). Stability's role as middleman has been so far innocuous if not generally helpful, but now that this role has a market price of $1B, companies and researchers are going to be eager to cut them out of the process and reap the rewards for themselves.
Runway is not "any company" they were part of the research for the original stable-diffusion release and were still working with StabilityAI team on SD. There must have been internal disagreements on the release of 1.5 version
This is really strange, and is being handled very badly.
RunwayML are an official partner of Stability AI. Stability has been delaying the release of their 1.5 model for several weeks now, citing unspecified "legal concerns".
RunwayML has seemingly gone ahead and released it themselves anyway, forcing Stability to file a takedown request for leaking their IP.
In the aftermath of the AUTOMATIC1111 debacle last week, this doesn't reflect well on Stability's comms. Things seem a lot less open since investors got on board...
It's a little confusing, but as far as a understand it, Stable Diffusion was created by a collaboration between RunwayML and Compvis (aka Machine Vision and Learning research group at Ludwig Maximilian University of Munich) with Stability.AI funding the computing power for training and LAION.AI (also funded by Stability.AI) providing the dataset.
The first few releases of model and code have been done by Compvis, and this one by RunwayML. More than permitted by the license, this seems to have been released by the actual developers.
I've seen speculation that it has been released this way to provide distance between Stability.AI from eventual future litigations, but it feels more like an internal ultimatum deadline.
> I've seen speculation that it has been released this way to provide distance between Stability.AI from eventual future litigations, but it feels more like an internal ultimatum deadline.
It could be both. A company it is not necessarily a monolithic entity.
Stable Diffusion was released with Open RAIL License which allows anyone to fine-tune and release their own models while complying the ToS of the original SD release. It's quite strange IP Leak is quoted in the take down request.
Pedantically iiuc, Stable Diffusion 1.3 and 1.4 were released with that license but 1.5 has yet to be released officially. Besides a verbal promise from Emad which isn't legally binding, there's no reason a) 1.5 ever has to be released, but b) if it is, since it's Stability AI's IP, they're free to put whatever license they want on 1.5 including a proprietary for-money license which having a public torrent leak cuts in the revenue from.
I get the impression we're about to see an ocean of litigation around AI and IP. If you're not a participant it will probably make good popcorn fodder. Though I don't doubt that there is a threat to graphics design and other creative jobs, may not be time to switch careers quite yet.
Because there is a lot of money and fear involved I agree that there will be a lot of litigation, but the outcomes will be mixed at best for the copyright holders.
Lots of people have memorized lyrics to copyrighted songs but no one is suing them for the memorization. They are only suing when the memorizer is trying to make money off of the song without permission of the copyright holder. The burden of proof is on the copyright holder to prove that the builder of the model was intending to distribute IP that they did not own.
A lot of people have sampled literally everything in music and it is very very uncommon to get sued for creating something substantially different. I fail to see how AI is any different.
People need to realize that IP enforcement is a very leaky sieve at best, and that the very concept of intellectual property is an affront to the nature of information and ideas, particularly in a digital realm where bits are free to be copied at will
The sooner we come to this realization, and those that stand to profit from IP enforcement bullshit take their monetization elsewhere, the better off we will be as a species.
It's premature to claim these "AI" models are equivalent to a human memorizing lyrics to copyrighted songs. They're much closer to XOR'ing copyrighted works so that the encoded form is not casually recognizable. I don't have sources close to hand, but I seem to recall during the 00's reading about such schemes being regularly proposed and also regularly struck down, legally.
From a high level both things consist of putting copyrighted bits in a digital blender, producing a mixture unlike the originals but which, if you prod it the right way, can be induced to reproduce some parts of the (copyrighted) input. I'm not sure calling it AI makes it (legally) different. What is the argument here - that previous schemes just didn't blend the bits enough?
> They are only suing when the memorizer is trying to make money off of the song without permission of the copyright holder.
Granted everything I know about copyright law I learned in a few audio tapes in my car, but I think the question I have is not, can you sue AI for absorbing information, but is AI unique enough that you couldn't find byte strings or patterns that match source material too closely.
This plays strangely with the: "Stability legal team reached out to Hugging Face reverting the initial takedown request, therefore we closed this thread" that is on the Hugging Face page.
Does anyone have a good resource on Stability, their history, and relationship to Runway? I haven't followed the space closely in a couple of years but they seemingly exploded from nowhere and I'm not entirely sure what their mandate is.
Readit News
> Cris here - the CEO and Co-founder of Runway. Since our founding in 2018, we’ve been on a mission to empower anyone to create the impossible. So, we’re excited to share this newest version of Stable Diffusion so that we can continue delivering on our mission.
> This version of Stable Diffusion is a continuation of the original High-Resolution Image Synthesis with Latent Diffusion Models work that we created and published (now more commonly referred to as Stable Diffusion). Stable Diffusion is an AI model developed by Patrick Esser from Runway and Robin Rombach from LMU Munich. The research and code behind Stable Diffusion was open-sourced last year. The model was released under the CreativeML Open RAIL M License.
> We confirm there has been no breach of IP as flagged and we thank Stability AI for the compute donation to retrain the original model.
https://huggingface.co/runwayml/stable-diffusion-v1-5/discus...
Edit: According to Discord messages from Emad Mostaque (founder of Stability AI) they have taken back the takedown request and seems this is the official release of the Stable Diffusion model version 1.5.
1.5 was apparently held back by Stability for weeks. Runway finally decided to just release it.
Stability requested a take down, and here you see the Runway CEO telling Stability in no uncertain terms "this ours to release, we created it, it's under an OS license, you don't hold any IP rights here; all you did was provide compute"
If anything this is a pretty stern rebuke of Stability and a sign of considerable disagreement between the two parties.
Deleted Comment
Deleted Comment
Deleted Comment
RunwayML are an official partner of Stability AI. Stability has been delaying the release of their 1.5 model for several weeks now, citing unspecified "legal concerns".
RunwayML has seemingly gone ahead and released it themselves anyway, forcing Stability to file a takedown request for leaking their IP.
In the aftermath of the AUTOMATIC1111 debacle last week, this doesn't reflect well on Stability's comms. Things seem a lot less open since investors got on board...
The first few releases of model and code have been done by Compvis, and this one by RunwayML. More than permitted by the license, this seems to have been released by the actual developers.
I've seen speculation that it has been released this way to provide distance between Stability.AI from eventual future litigations, but it feels more like an internal ultimatum deadline.
It could be both. A company it is not necessarily a monolithic entity.
https://huggingface.co/spaces/CompVis/stable-diffusion-licen...
https://opensource.org/osd
edit: never mind, jumped the gun and thought it was an official release but it turns out I was just confused.
https://huggingface.co/runwayml/stable-diffusion-v1-5
Lots of people have memorized lyrics to copyrighted songs but no one is suing them for the memorization. They are only suing when the memorizer is trying to make money off of the song without permission of the copyright holder. The burden of proof is on the copyright holder to prove that the builder of the model was intending to distribute IP that they did not own.
People need to realize that IP enforcement is a very leaky sieve at best, and that the very concept of intellectual property is an affront to the nature of information and ideas, particularly in a digital realm where bits are free to be copied at will
The sooner we come to this realization, and those that stand to profit from IP enforcement bullshit take their monetization elsewhere, the better off we will be as a species.
From a high level both things consist of putting copyrighted bits in a digital blender, producing a mixture unlike the originals but which, if you prod it the right way, can be induced to reproduce some parts of the (copyrighted) input. I'm not sure calling it AI makes it (legally) different. What is the argument here - that previous schemes just didn't blend the bits enough?
Granted everything I know about copyright law I learned in a few audio tapes in my car, but I think the question I have is not, can you sue AI for absorbing information, but is AI unique enough that you couldn't find byte strings or patterns that match source material too closely.
See the recent twitter thread on CSParse for example: https://twitter.com/docsparse/status/1581461734665367554