Monday, November 25, 2024
No menu items!
HomeTechnologyBefore Midjourney, there was NightCafe — and it's still kicking

Before Midjourney, there was NightCafe — and it’s still kicking

Elle Russell, co-founder of Cairns, Australia-based NightCafe, which offers a suite of AI-powered art-creating tools, prefers to avoid the spotlight.

“I like to remain hidden behind my monitors,” she told me in a recent interview.

NightCafe is similarly low profile.

The company, which Russell helped her partner, Angus Russell, launch five years ago, doesn’t get the same publicity as some of its rivals, like Midjourney. Yet NightCafe — an entirely bootstrapped venture that’s profitable “most months,” according to Elle — has enormous reach. Its over 25 million users have created nearly a billion images with its tools.

To pull back the curtain on one of the web’s oldest generative art marketplaces, I spoke with Elle about NightCafe’s origins, some of the challenges the platform faces, and where she and Angus see it evolving from here.

A website for wall art

As NightCafe’s founding story goes, Angus had recently moved into a semi-detached house in Sydney’s Inner West area and hadn’t had a chance to decorate it with much artwork. “You should get some art; the walls are bare,” remarked one guest. And while Angus agreed, he couldn’t find any prints online that spoke to him.

So in 2019, Angus, who had a degree in design and who’d co-founded a few design-focused startups, began a side hustle: a website where people could buy and sell AI-generated art. He called it NightCafe, after Vincent Van Gogh’s “The Night Café.”

It was an abject failure.

People liked creating the art, which NightCafe didn’t charge for. But they didn’t want to pay for wall prints, which was the only way the site made money.

Then one fateful week, Angus noticed that his hosting bill was a few hundred dollars higher than usual. Someone had generated thousands of images in just a few days. He implemented a credit system to prevent that from happening again.

Soon after, Angus’ inbox was flooded with requests to add an option to buy more credits, which he did. Practically overnight, the site became breakeven.

It’s at this point Elle joined NightCafe to run the business side of the operation. “I have two undergraduate bachelor degrees, in business and communications, and I’m also a CPA,” she said. “It made sense.”

NightCafe’s viral success

NightCafe got its second big break a couple years later, in mid-2021, when OpenAI announced DALL-E.

DALL-E, OpenAI’s first image-generating AI model, was state-of-the-art for the time. OpenAI opted not to release it, but it wasn’t long before enthusiasts managed to reverse-engineer some of the methods behind DALL-E and build open source models of their own.

Angus, who’d been closely following the developments, quickly worked to get one of the more popular DALL-E alternatives, VQGAN+CLIP, on NightCafe. He shelled out for hundreds of GPUs to scale it up.

The investment soon paid for itself.

Images created with NightCafe’s VQGAN+CLIP blew up on Reddit; NightCafe made $17,000 in a single day. Angus decided to quit his job at Atlassian to work on the platform full-time.

A model marketplace

The NightCafe of today is quite different from the NightCafe of several years ago.

The platform still runs some models on its own servers, including recent versions of Stable Diffusion and Ideogram. But it also integrates APIs from AI vendors that offer them, delivering what amounts to custom interfaces for third-party generators.

NightCafe
Selecting a model from NightCafe’s gallery.
Image Credits: NightCafe

That is to say, NightCafe layers tools on top of models from elsewhere, including OpenAI, Google and Black Forest Labs. And, as it has since 2019, the site provides printing services for customers who want mugs, T-shirts and prints of any art they generate.

“We’re a UI and community company,” Elle said. “NightCafe doesn’t have any internal AI or machine learning capability; we aggregate the available image models and make them fun and accessible to use.”

In NightCafe’s chatrooms, users can share their art and collaborate, or kick off “AI art challenges.” The platform also hosts official competitions where people can submit their creations for featured placement.

NightCafe
Chatrooms on NightCafe.
Image Credits: NightCafe

Last year, NightCafe introduced fine-tuning, which allows users to train a model to re-create a specific style, face or object by uploading example images. Fine-tuned models on NightCafe are subject to certain restrictions; for example, they can’t be trained on images showing nudity, celebrities or people under the age of 18, and they must be manually approved by NightCafe’s moderation team. (That’s to mitigate the risk of deepfakes.)

NightCafe
The terms users must agree to before submitting a fine-tuned model.
Image Credits: NightCafe

NightCafe is free to use, but only up to a certain number of images. Packs of image-generation credits can be purchased à la cart, and select features are gated behind a subscription. For fees ranging from $4.79 to $50 per month (undercutting Midjourney and Civitai), users get priority access to more-capable models, the ability to tip creators, the aforementioned fine-tuning capability and a higher image-generation limit.

It’s a model that’s worked exceptionally well for NightCafe.

A source close to the company tells TechCrunch that NightCafe is raking in $4 million in annualized revenue with a gross margin of nearly 50%, meaning that NightCafe is generating approximately $2 million a year in profit after expenses (inclusive of payroll for its nine staff).

Roughly a million people are visiting NightCafe each month, Elle says, and 20,000 have a subscription.

“Any AI art generator online is competing for money from the same people, though our users skew older than a lot of the industry,” she said. “We consider our biggest competitors to be other apps that have a strong community: Leonardo, Civitai and Midjourney.”

By opting not to train its own AI (and moderating fine-tuning), NightCafe is attempting to steer clear of the legal stand-off that’s ensnared many of the AI vendors whose models it aggregates.

Stability AI, Midjourney and a pair of other model providers, DeviantArt and Runway, face a class action lawsuit filed by artists who allege that the vendors engaged in copyright infringement by training their models on art without permission. (The vendors claim a fair use defense.) Some parts of the suit have been struck down. But a federal judge allowed it to move into the discovery stage early this month.

NightCafe may be protected by Section 230 of the Communications Decency Act, which holds users, not platforms, liable for illegal content (like copyright-violating artwork) so long as the platforms remove the content upon request. Australia, NightCafe’s home base, has the Broadcasting Services Act, which closely mirrors Section 230 with the exception that it imposes higher additional fees for failing to expeditiously remove “extreme violent material.”

Of course, should a court rule that the models NightCafe uses are essentially plagiarism machines, that’d be disruptive to the company’s business. But what about copyright as it pertains to NightCafe’s users and the art they generate?

NightCafe
Creating an image with NightCafe.
Image Credits: NightCafe

According to the platform’s terms of service, users retain the copyright for their AI-generated works in countries that recognize these types of works as copyrightable (like the U.S.) — at least as long as there’s permission to use any third-party branding, logos or trademarks within.

A post last May on NightCafe’s blog sheds more light on this: “Legitimate creators recognize and acknowledge where the inspiration used to create their images derived from another source. AI art creation tools are also evolving quickly, with systems in development to support the ongoing creative environment while ensuring that users can only access source material with the [consent] of the original artist — in much the same way that a royalty-free photography image may be permitted for use provided the creator is referenced.”

In other words, in NightCafe’s view, it’s the users, not NightCafe, who have to cover their bases. And if they don’t, the platform won’t defend them from the wrath of IP holders.

But it seems that IP holders don’t intimidate many users.

Cursory searches of NightCafe bring up images of Pokémon and Donald Duck, celebrities like Britney Spears, brands such as Coca-Cola and LEGO and artwork in the style of artists like Stanley “Artgerm” Lau. None appears to have been generated with the blessing of the copyright holders.

NightCafe
Image Credits: NightCafe

“Users can also report content that got through automated filters, and we have a team of human moderators working 24/7 on moderating flagged content,” Elle said when asked about this.

Political policies and deepfakes

As my interview with Elle segued to moderation, we dove into NightCafe’s general content guidelines, particularly its policies around politics and deepfakes.

Platforms, including Midjourney, have taken the step of banning users from generating images of political figures like Donald Trump and Kamala Harris leading up to the U.S. presidential election. But NightCafe hasn’t — and it doesn’t intend to, according to Elle.

“Generating images of Trump and other political and public figures is allowed,” she said. “However, we don’t want NightCafe to be a place for political arguments.”

How can NightCafe have it both ways? While the platform won’t prevent users from publishing political images elsewhere, it will flag those images for review if a user tries to post them to NightCafe’s public feeds.

That being the case, it’s trivial to find images of Biden in a wheelchair, Trump holding a gun and questionable Harris memes in NightCafe’s public gallery. With polls showing that the majority of Americans are concerned about the spread of AI propaganda and deepfakes, NightCafe certainly hasn’t made enforcement easier on itself.

NightCafe
Image Credits: NightCafe

As for what content is or isn’t allowed: It depends.

“Political bait,” glorification of divisive figures or purposely unflattering or demeaning images, are no-gos (in spite of what my searches turned up). Most content the average person would find harmful or offensive is also prohibited; NightCafe’s community standards list calls out things like racist and homophobic images, spam, offensive swear words, terrorism themes, images mocking people with disabilities, and depictions of hate groups and symbols.

These subjects may technically be disallowed. But type a term like “suicide bomber” into NightCafe’s search bar and there’s a decent chance you’ll come across at least one image that seems to fly in the face of the platform’s rules.

Elle tells me that it’s ultimately up to moderators to interpret NightCafe’s guidelines and that repeatedly publishing images in a banned category, or circumventing automated filters, could result in a warning or ban.

NightCafe has a rather small moderation team given its size (and the fact that the site’s users generate at least 700 images a day): five paid moderators and 20 volunteer moderators who get compensation in the form of premium NightCafe features. The paid moderators monitor content, while the volunteers handle comments, NightCafe’s chatrooms and the fine-tuned model queue.

Considering the poor working conditions content moderators are often subject to, I asked Elle for more information about NightCafe’s moderator recruitment practices. She said that the paid team is run through an outsourcing firm based in Indonesia (she wouldn’t name which) and overseen by an internal NightCafe staff member.

NightCafe
A few results for the search term “Coca-Cola.”
Image Credits: NightCafe

All paid moderators get a “market wage,” Elle said. (In Jakarta, the minimum wage was around $325 per month as of early 2024.)

Similar to Civitai, NightCafe has a policy carve-out for “NSFW” content: short of outright nudity, but permissive of suggestive poses (with “bare breasts and bums”), blood and gore, graphic depictions of war, and images of illegal drug use (e.g., Mickeys smoking blunts). This is somewhat dependent on the model; OpenAI’s DALL-E 2 has a stricter set of filters, for instance.

Why allow NSFW images despite the risks and without any form of watermarking (which might soon be legally mandated in California) to prevent abuse? To the first question, Elle says that it would stifle “artistic freedom.”

“We do allow mild artistic nudity and adult themes on the site when tagged as NSFW, but not outright porn. We’ve tried our best to ‘draw the line’ for our users in our community standards so that they understand what’s allowed and what’s not,” she added. “We pride ourselves on our community and being the ‘hub’ for all things AI art.”

From my few searches, NightCafe doesn’t seem overrun with boundary-crossing objectionable stuff. But I couldn’t help but notice that most of the “sexy” images featured women — an unfortunate pattern on platforms such as these.

Where NightCafe goes from here

Like many startups in the AI-powered art-generating space, NightCafe appears to be in a bit of a holding pattern. It’s bringing new models online, including video-generating models like Stable Video Diffusion. But it’s not rocking the boat too much — the unsaid reason being that a single court decision or regulation could force NightCafe to rethink its entire operation.

Still, Elle seems to think NightCafe has legs and doesn’t need outside investment.

“The majority of our competitors raised money over the last two years while image generation was hot,” Russell said. “Pretty much all of them were, or are, offering image generation at a loss to acquire users. Not all of them can succeed; NightCafe pioneered the intersection of AI and art but also championed the idea that creativity using advanced technology should be accessible for all.”

There’s no plans for an enterprise NightCafe offering, despite how lucrative such a product could prove to be (moderation roadblocks aside). Elle says that the focus will remain on building a community and “social hub” atop the latest generative models.

“One challenge that the industry faces is that image-generation models are getting so good, they’ll soon be commoditized,” she said. “What do companies compete on then? At NightCafe, we’ve chosen to focus on being an aggregator of the top models to provide the best variety and highest level of technology.”

We’ll see how it navigates the choppy waters from here.

RELATED ARTICLES

Most Popular

Recent Comments