Cool service! This, though, may prove to be a weakness of your business model. Having to sign up creates friction (I won't), which may lose you potential paying customers. What else could you be doing to give people a real feel for how well this works? Maybe a gallery of prompts and the respective results?
When I click to see an imagem from the gallery, it looks like it loaded but then a blank screen appears instead, after what looks like a refresh. Tried on Chrome and Firefox.
Not free, of course, but when I loaded Adobe Illustrator for the first time in a while the other day I noticed it had gained a feature where it can generatively add elements to vector illustrations. It worked very well in my casual playing around and maintained the style of the existing image too.
Are these generated by doing normal text to (pixel) image generation, then doing a raster to svg conversion? Or are LLM models actually capable of generating SVG directly?
I can imagine that working for simple things: make a red circle with a blue line under it, because there are many simple examples on the web.
But generating complex SVGs directly from text seems like it wouldn't work at all. The examples on their home page are well into the "too complex to generate by piecing together bits of svg examples seen during training."
Definitely raster to svg conversion. The giveaway (besides complex SVG generation being too difficult, as you say) is the splotchiness of the images, since the vectorization doesn't work super well with gradients.
With multimodal abilities, I wonder if it could eventually check a raster of the SVG it created and then make iterative improvements that get a better result than vectorizing a raster from stable diffusion.
Yeah unfortunately differentiable SVG rendering is still something that is very difficult to accomplish with diffusion models. I looked into this last year and found VectorFusion [0][1], which appears to be using a similar technique to this app - generate the raster image and then apply a vectorization tracer. This leads to the blobbiness you see, given the way the diffusion model generates the image and the tracer applies to the output.
There have been a few attempts (i.e. this recent one [2][3]) to attempt to fine-tune the parameters of the SVG with image segmentation, before comparing rendered outputs with a CLIP model. This is promising but the search space for vector images is just so huge that so far you really need to start with an existing image as a basis rather than starting from scratch. Interesting area of research!
Strange that the free tier has no documented limits. I always like it when I can see what I am working with, especially when refining an image will eat into that number.
The best pixel to vector is still vectormagic. They are on it since at least 2009 and have a native desktop app. I am not affiliated but just a bit flabbergasted that they are still so far ahead.
> By using the Services, you grant to VectorArt.ai, its successors, and assigns a perpetual, worldwide, non-exclusive, sublicensable, no-charge, royalty-free, irrevocable copyright license to use, copy, reproduce, process, adapt, modify, publish, transmit, prepare Derivative Works of, publicly display, publicly perform, sublicense, and/or distribute text prompts and images you input into the Services, or Assets produced by the Service at your direction.
-> even to images I upload? This is a no go. As a paid subscriber, I would expect to a) not automatically give a license for everything I UPLOAD
B) not automatically give a license for content I created myself
Yeah this is excessive in the extreme. Wave goodbye to your copyright if you upload images to this service.
The other thing is that they want you to pay to use the generated images commercially - but (as far as we know) purely AI generated images are not subject to copyright, so such licensing restrictions can’t exist.
> purely AI generated images are not subject to copyright, so such licensing restrictions can’t exist.
This doesn't feel right, similar to how it doesn't feel right that the companies are hoovering up copyrighted data. The company has invested in creating a cool product here, and it's reasonable for them to be able to license that work when used commercially.
They created a product and they definitely have copyright over that product.
However, what the product is creating, the output, is an AI generated Image... Which because it wasn't human authored, is automatically put into the public domain the moment it's created.
I really don't think this has been legally established, especially not in the case that AI is being used in conjunction with a copyright base image that it uses as a scaffold. It's incredibly hard to see the result as anything but a derived work of the original and being copyright by the author of the original in a similar way to applying a digital filter would.
The courts will sort out the boundaries of copyright here, but I'll eat my shorts if it's as extreme as the "all things that come out of AI are public domain" crowd thinks.
The gist of guidance from the US Copyright Office is that it's relative to what was put in.
If your input was a six word prompt of everyday langhage and the widely available magic box decompressed that into a cool picture, nobody produced anything copyrightable. And most of us wouldn't want that to be different. You'd just have squatters racing to claim copyright over every Midjourney prompt they can generate.
But if someone put what's clearly a demonstrably significant amount of commercial direction or artisanal effort into producing a original work while using generative tools, that direction/effort is what earns your copyright. The same would likely apply when using the tech to further "work" some already-copyrighted material of your own.
You're right that most of this stuff hasn't been tested by the courts when the tools are called "AI", but many of the underlying questions have decades and decades of precedent behind them (music sampling, collage and assemblage in visual art, monkeys taking snapshots, etc). Regulators do and lawyers will be making their cases about AI copyright in reference to those earlier, analogous, cases.
Yes, the “copyright base image that it uses as a scaffold” case is special and I agree that we’ve not seen the conclusion there yet. Though, it must be said that I’d expect to the extent that such copyright exists it’s going to belong to the original image author, not the model creator/user.
The images themselves may not be subject to copyright, but you can still enter into a contract that says you won’t use generated images without paying for them.
You can’t because copyright law supersedes contract law and copyright law states that such rights are exclusive - thus contract law cannot be used to usurp or invent a copying restriction. Licensing is the exclusive right of the copyright owner and when no copyright exists there can be no license (or de facto equivalent).
It’s not though, it’s a restriction on using the material produced by the service. Any contract terms regarding such are superseded by copyright law - the only rights belonging to the owner are those granted by copyright law and the only restrictions that may be placed are those that copyright law permits.
No, the rights are not superseded by copyright law. Why do you think that? Your agreement with the service provider is just like any other contractual agreement. Should you use one of the generated images (that you generated and used outside the terms of your agreement), you would not be sued for copyright infringement. You would be sued for breach of contract.
What you're saying is akin to signing an NDA, violating it, and then saying your First Amendment right supersedes the NDA. It doesn't.
It’s been federal law since 1976 that copyright “preempts” any other laws granting overlapping claims. This includes contracts, which fall under state law. Now, strictly speaking the courts are divided on how to treat the specifics of these issues with respect to EULAs, but the Genius vs Google case should leave one weary:
Haha, I wonder if we’re ever going to see a litigation over a EULA, where the service provider is trying to nullify some duty based on the fact they used AI to write the agreement which swapped Provider and User in a sentence.
How would VectorArt.ai even believe that their user has the rights to do so? If (for example) I upload some stock photo off the Internet, I absolutely do not have the right to grant such a license.
Yet this is essentially the same kind of license Facebook has when you upload any image to Facebook. And billions of people still upload content to them every day.
When asked what regulations Google is considering when it comes to bringing Bard to Canada, Sebastian cited the Online News Act, which would force tech giants and other companies to pay Canadian news producers for links posted to their platforms
“We’ll bring (Bard) here to Canada very soon as we manage through all the regulatory issues,” Sebastian said.
I think "sign-up pages" here needs some clarification from the mods (maybe referring to wait lists?). If everything that requires sign-up is banned then why this section further down:
> Please make it easy for users to try your thing out, ideally without barriers such as signups or emails. You'll get more feedback that way.
So, if anyone has a weekend to spare, setting up a competitor by running stable diffusion with a LORA for vectorized output with vtracer on the image output... should about do it. Seems at least one of these services pop up every day, and it's usually the same thing wrapped in different packaging.
It's interesting to see the prompts on specific examples, like https://vectorart.ai/3e5ead1b-60be-40f8-ae6f-25e14ca4e9a0
I guess "vectorized" is doing some heavy lifting there for converting to SVG. It's always funny to me to see random text generated, especially when it's either close to right (like this example) or just mangled shapes that were derived from the Roman or Cyrillic alphabet, but which still have some flavor of the original font, just with zero meaning.
So, this is just traditional image vectorization on top of existing AI-based bitmap image generation. It would be more interesting if this was actually AI-based vector image generation.
They also don’t say if they are self-hosting the image generation or rely on an external AI service, which may have an impact on data protection.
Thanks everyone for your feedback and support. I've been adding GPUs as fast as I can to meet demand and some of the earlier performance issues should be much improved.
69 comments
[ 0.22 ms ] story [ 305 ms ] threadThe style transfer is really impressive (watch past the underwhelming first example of the airplane).
I can imagine that working for simple things: make a red circle with a blue line under it, because there are many simple examples on the web.
But generating complex SVGs directly from text seems like it wouldn't work at all. The examples on their home page are well into the "too complex to generate by piecing together bits of svg examples seen during training."
I once asked it to create a SVG of Earth and it create a green ellipse on top of a blue circle.
There have been a few attempts (i.e. this recent one [2][3]) to attempt to fine-tune the parameters of the SVG with image segmentation, before comparing rendered outputs with a CLIP model. This is promising but the search space for vector images is just so huge that so far you really need to start with an existing image as a basis rather than starting from scratch. Interesting area of research!
0. https://ajayj.com/vectorfusion/ 1. https://arxiv.org/abs/2211.11319 2. https://intchous.github.io/SVGCustomization/ 3. https://arxiv.org/abs/2309.12302
https://vectormagic.com/
Not affiliated, but I've used them for a project and I was impressed
-> even to images I upload? This is a no go. As a paid subscriber, I would expect to a) not automatically give a license for everything I UPLOAD B) not automatically give a license for content I created myself
The other thing is that they want you to pay to use the generated images commercially - but (as far as we know) purely AI generated images are not subject to copyright, so such licensing restrictions can’t exist.
This doesn't feel right, similar to how it doesn't feel right that the companies are hoovering up copyrighted data. The company has invested in creating a cool product here, and it's reasonable for them to be able to license that work when used commercially.
However, what the product is creating, the output, is an AI generated Image... Which because it wasn't human authored, is automatically put into the public domain the moment it's created.
The product is copyrighted, the output is not.
The courts will sort out the boundaries of copyright here, but I'll eat my shorts if it's as extreme as the "all things that come out of AI are public domain" crowd thinks.
If your input was a six word prompt of everyday langhage and the widely available magic box decompressed that into a cool picture, nobody produced anything copyrightable. And most of us wouldn't want that to be different. You'd just have squatters racing to claim copyright over every Midjourney prompt they can generate.
But if someone put what's clearly a demonstrably significant amount of commercial direction or artisanal effort into producing a original work while using generative tools, that direction/effort is what earns your copyright. The same would likely apply when using the tech to further "work" some already-copyrighted material of your own.
You're right that most of this stuff hasn't been tested by the courts when the tools are called "AI", but many of the underlying questions have decades and decades of precedent behind them (music sampling, collage and assemblage in visual art, monkeys taking snapshots, etc). Regulators do and lawyers will be making their cases about AI copyright in reference to those earlier, analogous, cases.
What you're saying is akin to signing an NDA, violating it, and then saying your First Amendment right supersedes the NDA. It doesn't.
https://patentlyo.com/patent/2022/08/copyright-preempting-co...
Great era to be EULA hunting!
When asked what regulations Google is considering when it comes to bringing Bard to Canada, Sebastian cited the Online News Act, which would force tech giants and other companies to pay Canadian news producers for links posted to their platforms
“We’ll bring (Bard) here to Canada very soon as we manage through all the regulatory issues,” Sebastian said.
https://www.bnnbloomberg.ca/google-exec-says-regulatory-conc...
https://news.ycombinator.com/showhn.html
> Please make it easy for users to try your thing out, ideally without barriers such as signups or emails. You'll get more feedback that way.
The quote is them just asking nicely and clearly to make it accessible.
it can perceive what is being requested but you need to break down each part in a g statement just like with coding
They also don’t say if they are self-hosting the image generation or rely on an external AI service, which may have an impact on data protection.
https://Ebank.nz
Would be interested in some research to see if we can re use something I've done in the pixel domain