After AI chat comes the AI canvas. We put Lovart, a $2B+ design AI, to work

No film shoot. No models. No camera crew. We made this 15-second ad with AI.

The product is made up, too. It started with a passing thought: wouldn't it be nice if this drink existed? From there came a brand name, a bottle design, three flavors and, finally, this ad. All of it was made on a single AI screen.

We made a few other things as well.

On the left, a three-view sheet for a VTuber (a streamer who appears as an animated character), with front, side and back matching down to the hair and costume. On the right, a movie-style poster for DATA WORLD. Both came out almost on the first try from a chat request.

The tool is the design AI Lovart.

In Japan, Lovart is still something of an insider's secret. Globally, its growth has been remarkable.

Even in pre-launch testing, it had 800,000 users in 70 countries. It opened to the public in July 2025 and passed 10 million users by December, with annual recurring revenue (ARR) of $80 million (about ¥11.5 billion). Its website now says it's trusted by 30 million marketers, business owners and creators worldwide.

The money is on another scale, too. In June 2026, Lovart's developer, Evoken, raised about $300 million (about ¥49 billion) from investors including Tencent (21st Century Business Herald, 36Kr Japan). In August, Bloomberg reported early talks on a Hong Kong listing (Guancha).

Behind Lovart is Chen Mian, who led monetization for the global rollout of CapCut, the video-editing app from TikTok owner ByteDance (Fortune China). He helped put video editing on everyone's phone. His next bet was design.

WHAT IS

Lovart

An AI design agent that grows images, video and references on one canvas as you talk it through in chat

“The world's first design agent”── Lovart's description of itself on its official website

1. 2025.07Public launch800,000 users in 70 countries during testing. Debuts a combined chat-and-canvas screen
2. 2025.11Arrives in JapanFull Japan launch announced, with support for Google's Nano Banana Pro
3. 2025.12Passes 10 million usersAnnual revenue run rate reaches $80 million
4. 2026.06Raises about $300 millionValuation tops $2 billion
5. NOWOfficial hands-on workshop in JapanAugust 2026, with ByteDance's research team on stage. Major updates keep coming

Tell the chat something like “I want visuals for this kind of product.” The AI comes up with ideas, makes images and video, and lays them out on a wide canvas that keeps growing sideways. Compare ideas, fix one part, turn a still into video, export a PSD (a layered Photoshop file): planning through finishing, all on one screen.

It bundles top image and video models in one place: OpenAI's GPT Image 2, Google's Nano Banana Pro and Veo 3.1, ByteDance's Seedance and Seedream, Midjourney and more. Lovart picks the right one for the job, so you don't need to know which AI is good at what (Lovart).

EDITOR'S VERDICT

The editor-in-chief's verdict after four projects and a workshop

The whole planning meeting fits on one canvas.

No tool I've used gets from idea to something I can show and discuss this fast. Running the whole workflow at the workshop, I caught myself thinking Lovart might replace the creative production process itself.

Who will get the most out of it

• People who want to turn a product or campaign idea into something they can show the same day
• Teams that make many kinds of assets for one product, such as banners, social posts and videos
• People who like developing ideas while collecting reference images

I've been going to the gym lately, and it hit me: wouldn't it be handy to have one drink with hydration, protein and vitamins, the kind you'd grab right before a workout? That really was the whole starting point.

In chat, I gave the agent three things: the brand name “ReFUEL,” the slogan “Make your everyday workout extreme,” and a target of fitness fans in their 20s and 30s. Not much to go on. (The prompt was in Japanese, hence the Japanese copy on the designs below.)

Lovart chose an image AI on its own (GPT Image 2, this time) and laid out three concepts in different directions on the canvas.

Blue lightning. White and orange. Black and neon green. Each goes beyond a mockup with a name slapped on; you can already see who it's for and what impression it should make.

I picked the neon green, and Lovart immediately expanded it into the bottle's front and back, plus three flavors: Green Apple, Lemon Lime and Berry Boost.

That took about 20 minutes. A “wouldn't it be nice” idea in my head became something I could bring to a meeting. Ingredients, legally required labeling and print-ready files would still need separate work, but as a first step in product planning, it's more than enough.

Next, an ad. For video, I chose ByteDance's Seedance 2.5, announced only at the end of July and already selectable in Lovart.

I gave it the product image and a written direction (in Japanese): the product shines in sports scenes. Back came a video that flows from product close-ups to wide shots of the game and athletes in motion.

Then I added one reference image with the colors of an international brand's ad and the soft background blur of a cinema camera. The result is the video at the top of this article. It wouldn't look out of place in my Instagram feed.

Lovart publishes “skills,” packaged workflows for specific production tasks. Pick one by purpose (branding, advertising, social media, e-commerce, video), give it a short request and some references, and that's it.

My wife once mentioned she'd like to try being a VTuber. So I opened the character sheet skill and asked for “sakura pink, twin tails, a game-streaming VTuber, a cheerful little-sister type, a Japanese-style outfit.” (The prompt was in Japanese; the “little sister” type is a common anime archetype.)

To my surprise, it produced a three-view sheet that was exactly what I wanted, on the first try. A stray game-controller-shaped ornament disappeared cleanly with one pass of the eraser tool (left: first output; right: after touch-ups).

Lovart can also export multiple assets together as a single PSD file, and Inochi2D, software for animating 2D models, can import PSD files, so there's a path all the way to a VTuber model that actually moves.

Plenty of people collect references on Pinterest. I'm a heavy user, pinning constantly for work and hobbies.

Lovart connects to Pinterest and shows your boards on screen. Select images, click “Add to chat,” and they drop straight into your conversation as references.

I picked posters with photos cut into giant letters, a paper texture and slightly faded colors, then asked it to “make a poster for DATA WORLD.” The result is the poster at the top of this article. (The request was in Japanese. The line across the middle reads, roughly, “The world is more interesting when you look at it through data.”)

It came out genuinely stylish. When the place you collect references is connected to the place you make things, you can give an idea shape before it cools off.

On August 28, 2026, I attended Lovart's official hands-on workshop in Japan. Beyond the basics, it packed in business case studies and a look at the newest models from Seed, the ByteDance research team behind Seedance and Seedream.

In the hands-on part, Takahashi of Ouka Ichimon Inc. taught the basics while participants ran Lovart on their own laptops. We practiced handing over references and instructions, and fixing results that missed. Lovart also handed out free credits to everyone, a generous touch.

I'd already put plenty of hours into Lovart. But some creators and designers in the room were trying it for the first time, and you could see the novelty on their faces as they watched their screens.

Seedance 2.5 moves to 30-second single takes

ByteDance's Seed team walked through Seedance 2.5's progress. Single-generation length has doubled from 15 to 30 seconds, and you can now fix just a chosen stretch of time while keeping everything else intact (ByteDance Seed).

They compared the new and old models on light reflections, camera work, facial expressions and shattering glass, and screened a generated clip of friends at a Japanese izakaya (a casual pub) working out how to split the bill.

The highlight was a trailer for a fictional movie made with Seedance 2.5. My honest reaction: no filming, no CG artists, and it looks this good?

In e-commerce, fixing takes longer than making

What stuck with me most was a talk from a company that produces content for e-commerce (online retail).

VOICE

In e-commerce, we change the banner copy for every sale, and we swap product backgrounds and even people's facial expressions. When a logo changes, we fix past materials too. We spend more time fixing what already exists than making anything new.

For this they use Edit Elements, which splits an existing image into elements such as background, people and text so you can fix each separately (Lovart documentation). You can export a layered PSD, hand it off to Photoshop, and spend the time you saved pitching one more idea. Speed gets reinvested in the next piece of value.

They also shared how a young employee made several versions of a poster about the company's mission. Being able to make ideas visible and line them up, early in your career, is a big deal.

And it keeps evolving

The workshop also previewed a major update for the following month, pitched as letting creators hand off whole jobs, as if the AI finishes the work while you sleep. An app for making long-form video easily, no special skills needed, is also on the way (both per explanations at the venue).

I can't wait to try it.

Across all four projects, what helped most was how easy the canvas is to take in at a glance. ReFUEL's three concepts, side by side, practically invited a discussion of which works best.

In a chat AI, ideas scroll top to bottom, and comparing means scrolling back up. On a canvas, the ideas you didn't pick stay in view while you choose the next direction. It feels like pinning ideas to a whiteboard in a planning meeting.

And Lovart doesn't start drawing right away. It first plans the production, then picks a suitable AI for each task and writes separate instructions for each (Lovart, “MCoT Engine”). All I said was “I want to make this kind of product”; Lovart chose the image AI itself. Even without much AI knowledge, you can compete as long as you have an eye for good ideas. I think that's why it has reached more than 10 million people.

Lovart doesn't develop its own image or video AI. It bundles models from OpenAI, Google, ByteDance and others in one place.

For users, that matters. Seedance 2.5 was selectable in Lovart shortly after ByteDance announced it. No new sign-up every time a new AI comes out: you can tap the latest capabilities without leaving a screen you already know.

Founder Chen Mian led CapCut's global rollout at ByteDance. Seeing ByteDance's research team share the stage at the workshop, I sensed that this closeness, getting the newest models to creators early, is another of Lovart's strengths.

Relying on other companies' AI does leave Lovart exposed to their policy and price changes. Still, for users, not having to chase “which AI is best right now” is a big relief.

Some creators and designers at the workshop were trying Lovart for the first time. More than 10 million people use it worldwide, yet in Japan it's just getting started: Lovart announced its Japan launch in November 2025, less than a year ago (PR TIMES).

Flip that around, and anyone who starts now has a real first-mover advantage. Fixing e-commerce banners, drafting a first-pass plan, preparing many variations of social assets: every use I heard about at the workshop is something a company in Japan could try tomorrow.

Start by putting one “wouldn't it be nice” idea onto the canvas. ReFUEL took shape in about 20 minutes from an idea sparked by my gym routine.