What I like about Stable Diffusion is the tooling that's being developed around it. There is a "Comfy UI" which I've found to be too difficult, but there is one by Automatic1111 which is very accessible to me (the masses). That gives it a lot of potential in terms of ideas and what can be achieved. ControlNet adds a whole dimension to the generation process.
I am not familiar but is there something similar to ControlNet (at least conceptually) in MidJourney/DallE?
I can't keep up with the cutting edge between day job and kids. My ideas are either irrelevant or already done by the time I have time. I have SD and Llama 2 running locally, but at this point it's because I can, but any serious work I'd use a managed model to avoid maintenance and stay on current.
The New York comic example is pretty impressive. I tried creating a short comic story with both Dalle-2 and SD. After a lot of hacking I finally gave up an decided the technology was nowhere near ready. I may try again with this. I wonder how consistent it can be with its output. Can it generate a bunch of consistent and coherent comic book pages?
Fascinating, even poetic, that DALLE-3 imagined a simple vista with a sign that reads "Error establishing a database connection." It's almost as if AIs imagine a dystopian future where humans are cut off from the very networks they created.
>Some key advantages of DALL-E 3:
...
>Creative Freedom – Ability to generate any theme, style or subject matter with impressive coherence. No restrictive content moderation.
Did they remove censorship from this one? My understanding is that the previous versions were heavily censored.
The “cute white tiger cub smiling, with tail, sticker, transparent background” is pretty funny: the background isn't actually transparent, but it instead (imperfectly) reproduces the grey-and-white tiling that is usually displayed by image viewers when opening transparent backgrounds.
I wonder if this is significantly influenced by the images you find when doing an image search for "<search term> transparent".
A lot of sketchy clip-art sites insert the checkered background, probably to both symbolize transparency and also to prevent people from just right-clicking "Save as...". You want the user to search the download button on a page full of ads and fake buttons after all...
I think people that just enjoy drawing and art because it's "fun and creative" can still enjoy it even with AI. But if you want to be "the best", computers will indeed be a tough competition...
My prediction: Either these models will have to expose some editable intermediate step, or they have to become as smart as a human.
From the users point of view, a sentence is turned into an image. But there is an underlying structure to the pixels that we aren't allowed to tinker with.
Many choices are made like, placement, color pallete, level of detail, emotion evoked, or sub image descriptions (What should this bush look like, what kind of cloud should this be?)
These image generators are so useful because they fill in all the details you leave out, but you have no ability to tinker with the intermediate details they choose, and you can't just give them a paragraph of details either.
All of those things are already enabled for these models depending on the UI you're using to accessing. Discord even added a special inpaint UI to their application for Midjourney
There are models available that give you more control - in some senses, at least.
For example, you can use Stable Diffusion with 'ControlNet' [1] where for example, you can input an 'openpose' to choose the pose of people in the scene.
There's also a 'Regional Prompter' [2] which lets you use different prompts for different areas of the image, giving you some control over the composition.
You can also use 'inpainting' to regenerate select parts of your image if, for example, you don't like the shape of the clouds.
Of course this stuff isn't perfect - for example, you'll get hands with the wrong number of fingers sometimes, no matter what you specify. And you can't easily generate things like multi-frame cartoons without characters clothes changing between frames.
From what I understand about DALL-E 3 is that it integrates with ChatGPT so you can ask it to change the image with simple language.
So if you generated, say, a picture that ended up including a bush, you could ask it to change it from a boxwood to a rose bush, and it could do it.
As it is, MidJourney already has the the ability for you to select an area of a generated image and make it regenerate it. You can even change the prompt, though it won't always do what you think.
I think a lot of the challenges you're describing won't apply to DALL-E 3's ChatGPT integration.
My understanding is that DALL-E 3 is supposed to reject attempts to mimic living artists. That would explain why the "Gary Larson" image doesn't resemble his style, but I'm surprised to see Spongebob being so on-model.
The improvements from previous models are impressive, but there still seems to be a lot of trouble properly matching actions to figures. I'm regularly seeing speech bubbles coming from the wrong characters (see the "blueberry" cartoon in this example). Overall, the best output still seems to be in the general realm of commercial art -- logos, simple figures, stock photo type images. I'm not seeing any huge improvements for narrative forms like comics.
If copyright law doesn’t take a massive leap forward in the next year, AI is going throw a wrench in our entire economy. The creative humans that trained these models deserve attribution, compensation, and royalties.
The ongoing AI era is reminding me a lot of early Uber/AirBnB - these companies grew faster than local laws could keep up, and now we have millions of gig workers without insurance and housing crises in most major cities. We need to hit the breaks on this tech so it can be deployed safely.
How is THC being 'legally' sold even in states where it is illegal? Government in the US has proven too inefficient to handle the speed of marketplace innovation. You can't outlaw ingenuity.
I don’t think there’s any value in comparing copyright with American drug laws. They’re not remotely similar and drug law in America is one of the most politically controversial arenas one can enter.
A much more fair comparison would be existing copyright law, which provides avenues for ownership without the ability to create new derivative works. It’s possible to expand these laws.
on the contrary, it's proof that popular support for something can overcome even what was a fairly brutal crackdown by large men with large guns and quasi-military strength. people like dalle and midjourney and chatgpt.
So what? Weed has had popular support for 60 years… You’re giving an example of a law that has very slowly changed across America and arguing that laws are very slow to change. It’s just irrelevant to this conversation. You’re talking about the supply/demand of pot while I’m talking about the economic consequences of copying human work without compensation. I agree ChatGPT is popular. So what?
Certain refrigerants and aerosols were very popular. Then we found they were blowing a hole in the ozone layer, so they were regulated. Popular things are sometimes regulated.
As hardware progresses it will become increasingly possible to train your own dragon in your own catacomb. I'm unsure what kind of edge it would give to do an unethical/illegal employee review, credit rating, investment portfolio etc etc but if it only seems like it would give an advantage people will do it.
Copyright law isn't going to matter here. There are already models trained solely on owned or licensed content and there is more than enough funding to produce more. The big businesses dumping billions of dollars into AI have zero interest in producing counterfeit Calvin and Hobbes strips.
I don't know that I personally know how we get to the ideal outcome here, but I see a lot of ways in which a copyright crackdown will have unexpected ramifications (accidentally banning search indexes, etc.) and consolidate power in the emerging market even further into the hands of extremely well-funded incumbents.
> The big businesses … have zero interest in producing counterfeit Calvin and Hobbes strips
Agree. I’m not suggesting creative authors are attributed and receive royalties based on work generated in their style. I’m suggesting they get compensated for work generated from their training data - based on model weights.
People need to get paid. Again, this is Uber/AirBnB all over again except it applies to literally every industry.
I don't think the question of rights conferred through training data is going to be resolved through copyright law. I think the availability of models trained exclusively on licensed data is going to bypass that controversy. People will get paid to contribute data but probably not a lot.
Sure, and that licensed work was created by people. Those people should get paid upon the creation of new derivative work. That’s exactly what copyright law is.
Okay… I feel you understand what I’m getting at but are super keen on nitpicking. I know copyright doesn’t mandate payment schedules. It’s a legal tool that prevents people from distributing derivative work without permission. Currently (most) AI models don’t attribute original creators (let alone pay them royalties), which means it’s impossible to pay content creators even if we wanted to.
It’s a widely known issue that’s a really hot topic right now.
Do you have a sense of how much people tend to get paid for commercial art, stock photography, etc.? That's essentially the ceiling on what contributors to training datasets can expect going forward. Regulation isn't likely to be able to significantly influence that, at least not in the US.
I do. The royalty system for music, TV, and movies are all actually quite mature - artists receive a few pennies per play. It adds up quick and can easily make a living on royalties.
And why would that be the ceiling? AI only expands their existing markets and revenue streams
Royalties are almost 100% a private industry matter in the US. There is compulsory licensing with legally mandated royalties for certain types of musical performance, but that is used quite rarely. The vast majority of US music licensing goes through performance rights organizations like ASCAP and BMI.
I also wouldn't characterize musicians as being happy with the current state of royalties. Ask a moderately successful musician how much they make on Spotify plays sometime.
> The creative humans that trained these models deserve attribution, compensation, and royalties.
I would much rather give them nothing. They can participate in the system where we work for a living. Do useful things. People think everything is valuable. The Romans, in the end thought only the circus was important. Turns out you cant wear circus, you cant eat it, the circus makes non of the things we need.
If you could chose what to spend the worlds attention on. Would you pick the circus?
People will continue making things regardless, for fun and contract work.
I know it is the norm but I don't want all 5000 possible variations of the letter E owned by someone.
Trademarks we can keep, those are actually useful.
By definition, a royalty is the sharing of already realized value. We don’t need to imagine what value may be realized, so the Roman Circus analogy isn’t necessary or accurate.
There is a difference between the [collective] perception of value and what is objectively valuable for our civilization.
We might imagine an island with 100 inhabitants where 3 people own 99% of the wealth and 99% of the produce. Here valuable is for the most part defined by what those 3 people value. If they really enjoy good choreography, 50 people could permanently work to put great dance shows together. It makes senses to cut down on fishing and farming. If there is food for 40% the system works.
China deleted 12,000 cryptocurrency influencers accounts from Weibo and Baidu. Both cryptocurrency and influences are valuable. As the formula creates it's own money everyone could be a cryptocurrency influencer for a living. I don't see why the sector couldn't grow to be bigger than the rest of the economy.
To me the danger starts when wealth and income are increasingly controlled by people who don't make useful contributions to our collective survival.
We can use information to educate people and to do countless productive things.
I'm not willing to sacrifice 1% of that process to rent seeking organizations and individuals or to peoples sensitivities. We are not talking about 1%, we are attempting to design the entire flow of information to cater to these people primarily.
> There is a difference between the [collective] perception of value and what is objectively valuable for our civilization.
No, not perceived value. I’m literally talking about sharing revenue. Dollars. Cold hard cash.
Eg Customer pays to have DALL-E generate an image, then fractions of a penny are shared with contributors of the DALL-E model based on the contributors’ weights that were used.
Again the rest of your analogies are unnecessary because we’re talking about two different things.
can you explain why you think obsoleting the majority of commercial artists will "throw a wrench in our entire economy"? since you're addressing copyright law i'm assuming that's all you're addressing. this sounds like a very small number of people. for the economy as a whole, bringing down the cost of art is probably a positive.
i understand the arguments about artistic value, but those are quite separate from yours.
The vast majority of artists are working professionals doing the art grunt work, so it’s many more than you’d think (Millions) - not just those working for film/media/galleries. That said, these models won’t only replace artists. They can be trained to do most anything - jobs in marketing, legal, journalism, healthcare, software, etc are all at risk
right, this is why i'm specifically questioning your assertion on copyright law. you weren't talking about ai displacing jobs generally, just copyright law - which doesn't protect jobs in marketing, legal, journalism, healthcare, and probably not software. i don't understand why you believe that "if copyright law doesn't take a massive leap forward in the next year" this sort of thing will happen.
I see what you’re saying. What kind of law or regulation would cover the rehashing of original content then? Isn’t that copyright law? It’s my understanding that published works in these fields would be covered by copyright law.
Prompt: "illustration of orbital nuclear explosion, planet earth in the background, shockwave, epic, super nintendo style pixel art"
"Unsafe image content detected. Your image generations are not displayed because we detected unsafe content in the images based on our content policy. Please try creating again with another prompt."
63 comments
[ 2.8 ms ] story [ 206 ms ] threadI am not familiar but is there something similar to ControlNet (at least conceptually) in MidJourney/DallE?
The model was developed by OpenAI
Oh nevermind the site is down. We're okay!
That's how you earn the reputation for producing CSAM.
A lot of sketchy clip-art sites insert the checkered background, probably to both symbolize transparency and also to prevent people from just right-clicking "Save as...". You want the user to search the download button on a page full of ads and fake buttons after all...
And also, can you fetch the un-watermarked image like you can in the Dalle Playground?
I think people that just enjoy drawing and art because it's "fun and creative" can still enjoy it even with AI. But if you want to be "the best", computers will indeed be a tough competition...
From the users point of view, a sentence is turned into an image. But there is an underlying structure to the pixels that we aren't allowed to tinker with.
Many choices are made like, placement, color pallete, level of detail, emotion evoked, or sub image descriptions (What should this bush look like, what kind of cloud should this be?)
These image generators are so useful because they fill in all the details you leave out, but you have no ability to tinker with the intermediate details they choose, and you can't just give them a paragraph of details either.
For example, you can use Stable Diffusion with 'ControlNet' [1] where for example, you can input an 'openpose' to choose the pose of people in the scene.
There's also a 'Regional Prompter' [2] which lets you use different prompts for different areas of the image, giving you some control over the composition.
You can also use 'inpainting' to regenerate select parts of your image if, for example, you don't like the shape of the clouds.
Of course this stuff isn't perfect - for example, you'll get hands with the wrong number of fingers sometimes, no matter what you specify. And you can't easily generate things like multi-frame cartoons without characters clothes changing between frames.
[1] https://github.com/Mikubill/sd-webui-controlnet [2] https://github.com/hako-mikan/sd-webui-regional-prompter
So if you generated, say, a picture that ended up including a bush, you could ask it to change it from a boxwood to a rose bush, and it could do it.
As it is, MidJourney already has the the ability for you to select an area of a generated image and make it regenerate it. You can even change the prompt, though it won't always do what you think.
I think a lot of the challenges you're describing won't apply to DALL-E 3's ChatGPT integration.
The improvements from previous models are impressive, but there still seems to be a lot of trouble properly matching actions to figures. I'm regularly seeing speech bubbles coming from the wrong characters (see the "blueberry" cartoon in this example). Overall, the best output still seems to be in the general realm of commercial art -- logos, simple figures, stock photo type images. I'm not seeing any huge improvements for narrative forms like comics.
The ongoing AI era is reminding me a lot of early Uber/AirBnB - these companies grew faster than local laws could keep up, and now we have millions of gig workers without insurance and housing crises in most major cities. We need to hit the breaks on this tech so it can be deployed safely.
How so? Circumvention would be illegal. Sure there might be a black market, but you couldn’t scale a real business off it.
A much more fair comparison would be existing copyright law, which provides avenues for ownership without the ability to create new derivative works. It’s possible to expand these laws.
Certain refrigerants and aerosols were very popular. Then we found they were blowing a hole in the ozone layer, so they were regulated. Popular things are sometimes regulated.
I don't know that I personally know how we get to the ideal outcome here, but I see a lot of ways in which a copyright crackdown will have unexpected ramifications (accidentally banning search indexes, etc.) and consolidate power in the emerging market even further into the hands of extremely well-funded incumbents.
Agree. I’m not suggesting creative authors are attributed and receive royalties based on work generated in their style. I’m suggesting they get compensated for work generated from their training data - based on model weights.
People need to get paid. Again, this is Uber/AirBnB all over again except it applies to literally every industry.
Coincidentally, Business Insider wrote about this topic just a few hours ago. https://www.businessinsider.com/microsoft-jaron-lanier-ai-ad...
It’s a widely known issue that’s a really hot topic right now.
https://hbr.org/2023/04/generative-ai-has-an-intellectual-pr...
Any existing models will be irrelevant in 6 months. Regulations enacted today would provide a legal foundation for tomorrow’s models to abide to.
And why would that be the ceiling? AI only expands their existing markets and revenue streams
I also wouldn't characterize musicians as being happy with the current state of royalties. Ask a moderately successful musician how much they make on Spotify plays sometime.
Also I said in my comment it’s pennies per play. Ask a moderately successful writer how much they make on chatGPT. Oh wait it’s zero.
I would much rather give them nothing. They can participate in the system where we work for a living. Do useful things. People think everything is valuable. The Romans, in the end thought only the circus was important. Turns out you cant wear circus, you cant eat it, the circus makes non of the things we need.
If you could chose what to spend the worlds attention on. Would you pick the circus?
People will continue making things regardless, for fun and contract work.
I know it is the norm but I don't want all 5000 possible variations of the letter E owned by someone.
Trademarks we can keep, those are actually useful.
By definition, a royalty is the sharing of already realized value. We don’t need to imagine what value may be realized, so the Roman Circus analogy isn’t necessary or accurate.
We might imagine an island with 100 inhabitants where 3 people own 99% of the wealth and 99% of the produce. Here valuable is for the most part defined by what those 3 people value. If they really enjoy good choreography, 50 people could permanently work to put great dance shows together. It makes senses to cut down on fishing and farming. If there is food for 40% the system works.
China deleted 12,000 cryptocurrency influencers accounts from Weibo and Baidu. Both cryptocurrency and influences are valuable. As the formula creates it's own money everyone could be a cryptocurrency influencer for a living. I don't see why the sector couldn't grow to be bigger than the rest of the economy.
To me the danger starts when wealth and income are increasingly controlled by people who don't make useful contributions to our collective survival.
We can use information to educate people and to do countless productive things.
I'm not willing to sacrifice 1% of that process to rent seeking organizations and individuals or to peoples sensitivities. We are not talking about 1%, we are attempting to design the entire flow of information to cater to these people primarily.
No, not perceived value. I’m literally talking about sharing revenue. Dollars. Cold hard cash.
Eg Customer pays to have DALL-E generate an image, then fractions of a penny are shared with contributors of the DALL-E model based on the contributors’ weights that were used.
Again the rest of your analogies are unnecessary because we’re talking about two different things.
i understand the arguments about artistic value, but those are quite separate from yours.
The vast majority of artists are working professionals doing the art grunt work, so it’s many more than you’d think (Millions) - not just those working for film/media/galleries. That said, these models won’t only replace artists. They can be trained to do most anything - jobs in marketing, legal, journalism, healthcare, software, etc are all at risk
"Unsafe image content detected. Your image generations are not displayed because we detected unsafe content in the images based on our content policy. Please try creating again with another prompt."
Sigh, safety is so annoying.