Canva's AI Features Have Been Used 10 Billion Times. The Playbook Is Not a Chat Box

Nathan Labenz, host of *The Cognitive Revolution*, sits down with John Milinovich, head of generative AI product at Canva. The company now serves more than 200 million people with a mission to let anyone design anything. Milinovich's path is not the usual PM loop: he studied architecture, later worked content quality at Pinterest across a corpus of four billion pins, founded a company, and now owns Canva's GenAI product. Labenz, who runs Waymark making video ads for small businesses, has watched Canva for years as a narrower cousin of the same problem.

The first cut is philosophical. Creativity, in Milinovich's usage, is generating new ideas; design is goal-directed work that has to communicate something or finish a mission. Creativity is what people enjoy most. Design is more mixed. Canva's user research landed on a perfect three-way split: a third of people want to spend more time designing, a third want to spend less, and a third want about the same. The company's job, as he frames it, is to break down barriers to global creativity and give people tools wherever they sit on that spectrum. How much automation versus augmentation they want depends on whether the task feels like expression or like production.

Adobe, Labenz notes, already owns the specialists. Canva still tries to serve the mix inside those 200 million users. The gap that built the company, Milinovich says, is creative confidence: helping people who would not otherwise feel like designers express themselves in a way that makes them proud. That is a hard product problem and, he thinks, the work of the last decade. Labenz's canonical Waymark user is a baker who opened a shop because she loves dough and now has a pile of other jobs. Those people mostly wanted the thing done for them. "Clients don't want software, clients want service," a line he has remembered for years. Looking at Canva, he saw world-class local utilities, especially background removal, and less of a ten-word-in, whole-campaign-out generator. That fork is the rest of the hour.

At the highest level, Milinovich says, there is automation and there is augmentation. Automation removes the ick from something super annoying. You know the outcome; you want it done. Background removal is the clean example. So is Magic Switch, turning a document into a deck, and Magic Sort on whiteboards, taking a pile of sticky notes from a brainstorm and putting them in a new order. Augmentation is the opposite posture: you are in the driver's seat, working toward something you cannot quite see yet, and you want a suite of capabilities that turn a fuzzy feeling into something concrete. Outcomes there are fuzzier. You know it when you see it.

The portfolio is organized on those two axes, and the strategy for both is mostly embedded AI: put the model at the point of need inside the existing editor. A more separate, AI-native experience is newer. Dream Lab, built with the Leonardo team Canva acquired, is the first real point of view on that. Magic Media, which lives in the editor, helps you while you are already making a social post for a daughter's first birthday. Generation in general can serve a much wider set of jobs that are not tied to a specific design. On the creativity-versus-design line from the top of the show, Dream Lab sits closer to creativity, for people who have not yet decided what they are making. Labenz tried it live: an astronaut-dinosaur birthday invitation for his almost-six-year-old, complete with "Ernie's sixth birthday" set in a plausible font, plus some invented address-looking glyphs underneath that he erased. Original Stable Diffusion, he reminds himself, could not do any of that. Waymark still has not shipped a straight text-to-image generator, because small-business users need to represent a real shop, not astronaut dinosaurs; Labenz's stack stays higher-level, giving the model a structure and asking it to fill dimensions.

Melanie Perkins, Canva's CEO and co-founder, has a line that lives rent-free in Milinovich's head: Canva's original innovation was moving design from a world of pixels to a world of objects. Stock photos, graphics, video, fonts, and templates lived in one library; you dragged and dropped in an editor ordinary people could stand. That is what made creativity accessible. AI, he believes, is moving the stack again, from object-level manipulation into concept-level manipulation. You will always need objects. You can now also do higher-order operations: ask for the vibe of a logo instead of dragging a rectangle.

Claude's Artifacts product is his text-side analogy. Anthropic made a deliberate choice not to let you select and tweak a specific span; you ask for conceptual edits across the whole generated markdown file. Image generation is catching up through multimodal diffusion transformers, ComfyUI pipelines, and ControlNets: depth maps for architectural rendering, pose estimation, content models that keep semantic meaning, style control. Krea and Leonardo's real-time generation let you drop abstract rectangles and circles, add a prompt, and see what comes out. That, he thinks, is concept-level control starting to win at the image layer, similar to the old jump from pixels to objects.

Should the product expose concepts to the user while asking the model to work at objects, or at pixels? It depends, and he actually maps it. Canva supports more than 500 design categories, which you can lay on a spectrum from form to function. Logos and social posts sit toward form; docs and whiteboards sit toward function. Function-heavy jobs can start at the concept level, Claude-Artifacts style, but users still need to tweak the wording. Form-heavy jobs can start with "what's the vibe of this logo" and still need a color control at the end. Language models and diffusion models are converging in architecture and in experience. The hard industry problem is a native interface on top that is not confusing. Labenz sketches a future of "tell us what you want," a first draft, then drop into objects. Milinovich does not want to zag just because the industry zags toward an open text box.

People want more or less the same amount of design, he says, but they communicate differently, especially globally. Ask three professional designers to describe a design, or five to explain what modernism means for a logo, and you will not get one answer. Open-ended text got humans used to the new substrate because it is a natural interface for token-in, token-out models, not because it is the best UI for AI. Design is one of the most multimodal disciplines there is: image, touch, gesture, voice, text. If you show four options, typing "I like the third one" is worse than clicking it. Challenge the assumption that text is the ubiquitous AI interface, and put just enough UI at the right moment in the journey to capture context.

Labenz wants software that works in the background every day, sending a seasonally appropriate video even if people publish 10 percent of it. Milinovich's previous company, Aesthetic, was a tech-enabled design agency with a 2018 thesis: the world will need more design than there are designer co-founders, so give every company one. GANs and Ian Goodfellow's work made them believe AI could eat not just the back office of an agency but the core work, especially because agencies sell the work, not the hours. He still believes AI will eat into services. At Canva, creativity stays central. Agentic workflows are for production design, the stuff you can almost see in your head the moment you think of it. The point is to remove friction, not humans.

The old era required a human to learn every tool and chain several of them to finish a job. The new era, as he tells it, is AI adapting to the human and the job, sitting across the whole workflow, understanding a content calendar, making suggestions. None of that, he insists, comes at the expense of humans remaining core. Labenz sees very few videos downloaded from Waymark without some human edit. He reaches for *The Simpsons*: Mr. Burns at a gallery unveiling, "I'm no art critic but I know what I hate." Users say that ain't it, then that's close, then they drill into the part that bugs them. That itch is why object-level steerability is not optional.

The same lesson showed up at Pinterest. With four billion pins, new content had to be judged good enough to show anyone, and historically the only evidence of "good" was whether people engaged. After a lot of research, the team found it is quite easy for machine learning to predict that something is bad or low quality, and extremely difficult to predict that something is very, very high quality. Foundation-model multimodal systems have not repealed that. An AI can predict quality only so far; you still need human intervention. A fully automated brand posting on every channel is a cool art project and, he thinks, far from reality, because people want their taste on anything that will carry their name.

Claude's computer use made Labenz point one AI at Waymark; it was surprisingly good at prompting other AIs, shaky on fine UI. Should smaller companies keep building drill-down UIs for humans, or bet that verbal control will feel like ownership? Milinovich's intuition is that models will keep getting better at writing prompts for other models; he likes asking an LLM to prompt a diffusion model, or Claude to prompt Perplexity, and it beats what he would write. Practically, not every company will need API integrations to connect software. The mental model is closer to Linux pipes than to a super-app sitting on top. If you are building an AI-native experience, expose the model. Always let a user interact with the AI as an escape hatch, independent of the opinionated UI you ship for the masses. That is true for the human user and especially true for the AI user. What is interpretable for an AI, he says, is usually easiest for humans too.

Labenz sometimes opens Canva just for background removal, even when he needs nothing else around it. The Vienna team behind Kaleido, now a core of Canva's generative research group and still running remove.bg as a separate property, is why that tool leads the market. Canva's AI features had already been used over 10 billion times, a number Milinovich cites as proof that the embedded approach found breadth. Expedia used Magic Switch and Magic Resize to crush workflows that used to take eight hours into about 20 minutes. A real-estate investment trust he names FN got "crazy time leverage" from Magic Edit. His own daily drivers are Magic Sort, Magic Write's personalized tone of voice, and Magic Layouts: dump stuff on a page and get dozens of layouts that add form to a purely functional pile.

Labenz's own white whale is writing. For podcast intros he feeds Claude a PDF of 30 past essays plus the current transcript, sometimes around 50,000 tokens of prior voice, plus a few runtime instructions. Adjectives, or a single sample, do not nail it. Canva has evaluated that curve for Magic Write and found it is very brand- and person-specific; they have not been able to publish a prescription for how many tokens make Nathan's voice versus John's. More is generally better. Users can put in quite a bit of text, thousands of tokens if not 50,000. Milinovich has compared it to a Claude Project of prior posts plus a stream of consciousness; his qualitative read is that it works extremely well, a mix of prompt work and UX that guides people. It is also a clean example of concept-level editing of tone while leaving object-level control of the actual words.

The model strategy is three-pronged and dates to 2017: build their own (more than 100 models in production), partner with research labs for best-of-breed, and pull in an app-partner ecosystem natively. In the LLM realm there is so much movement that trying many approaches and pitting them against each other has paid off. Labenz asks the GPT-4o image-editing teaser question and the more general one: how do you invest knowing a lab capability you cannot match is probably coming? Milinovich's industry-level answer is that the GPT moment for diffusion was multimodal diffusion transformers, a single transformer with a serious text encoder and cross-attention that actually aligns text to image semantics. That is what unlocked infill, expansion, and a new generation of image tools, progress he puts on par with the text era. Canva gets to watch the industry and also train and launch its own models with the Leonardo team.

With 200 million users you cannot run on vibes, and taste has no objective ground truth. Chatbots have LMSYS-style arenas; design does not. For two years Canva has gone extremely deep. Milinovich says they had to go to the end of the earth and become the best in the world at this. The stack is rigorous custom guidelines that describe the human-taste part of "good," a trained set of professional designers as evaluators large enough for statistical significance, category-specific golden prompts (a good deck is not a good social post is not a good whiteboard), and multiple generators competing on five to ten dimensions such as visual quality and prompt adherence. All of it is in service of avoiding homogeneity and giving users steerability.

Labenz sketches a graduated launch: model-vs-model smoke tests, trusted humans, then live "pick one" in the product. Milinovich separates offline eval for generation from online eval, which looks more like engagement of A versus B. Success has to be defined first. For background generation, the rubric goes three layers down: was the right foreground item picked, was the crop good, and if so was it over-cropped or under-cropped. Pregenerate examples, show them to enough evaluators, get a ground-truth good/bad. Three humans on the same crop may not agree, which makes training a judge model hard. In a narrowly scoped task you can still reach 70, 80, even 85 or 90 percent agreement with a human evaluator. Open-ended, style-dependent work is much harder. After offline eval, they interleave a new generator in production against other versions and watch uplift, the same logic as recommendation or search, applied to generation.

Fine-tuning multimodal foundation models for those judge tasks still blows his mind compared with Pinterest, where they trained from scratch. Five thousand examples of background generation or social posts go much further now because the base model has already seen millions or billions of images. Rules of thumb depend on the model class. A LoRA on a diffusion model can lock a style with tens of examples. Brand tone in writing takes more, and LLMs overfit easily if you starve them of examples. Classification is a good fine-tune; capturing a vibe is usually not. His stronger advice is to challenge why you are fine-tuning at all. With million-token windows, dropping 30,000 to 50,000 tokens of prior writing is, nine times out of ten, a better yield. Per-account fine-tuning is not a big part of Canva's strategy today, except perhaps aesthetic style on the image side.

Cost and latency matter at 200 million users, and still should not run the early design. Canva's development flow is make it work, make it good, make it fast, then make it cheap. Launch milestones usually sit after "good," sometimes after they start making it fast. The team is cost-aware; cost is not the primary driver. Zoom out, he says, and inference cost curves have dropped on the order of 95 percent in the last year. Over-rotating on cost at the start is a mistake. Nail the experience. Labenz wants it on a T-shirt. They joke they will design the shirt in Canva.

Hiring follows the same non-purist bent. Canva's first, second, and third priority in generative AI is to be the best in the world at design AI. Passion for that mission is the door test; you do not have to be a designer, but you have to be excited about humans and AI in creative work. Then curiosity, growth mindset, living in a world that changes every day, and actually using Canva daily, which is part of the culture. They want practical problem-solvers who will wear multiple hats, ship a model into the product to feel its jagged edges, and not wait for perfect offline metrics. The habit of mind he names is an unwillingness to be blocked: raise your hand for help from another team or from business development instead of waiting, and grind through problems that literally have no industry precedent, from a user- and domain-centric first-principles view.

Labenz offers two of his own. Do the hard part first: you cannot plan an AI product end-to-end the way you plan traditional software, so get a walking skeleton from inputs to outputs and figure out plumbing later. Milinovich agrees; often the unknown is smoothing the jagged edge. The second is situational awareness: tenacity that grinds 79 to 82 percent can hide a need to reconceive the problem. Milinovich prefers a sharper frame, borrowed from Dave Hearnden, Canva's former CTO and now a distinguished engineer. Humans have had the idea for hoverboards for 50 years; nobody has ridden one. The outcome is obvious; delivery is extremely difficult. If a promising solution gets you from 72 to 81 and you need 98, keep the problem and switch the solution. That is hard, because you have to stay passionate and still take a sidestep.

For inspiration he watches Krea on the blend of concept and object, Meshy and Spline in text-to-3D, and, because he studied architecture, Spacemaker and Ark Design. Dusting off Rhino and AutoCAD against that generation of tools still blows his mind. Daily drivers are Perplexity, Claude, ChatGPT, and Otter. He first fell in love with generative design in 2008, in architecture school, listening to a lecture on Processing that introduced randomness into an algorithm and opened parametric design: knobs for rooms and square footage, variations without AI. The new layer is concept-level. You can ask for a 2,000-square-foot house on a site plan and get a rendering or even a building-information model. You can simulate indoor temperature from sun and GPS to place windows, or use predicted paths to bathrooms and cafeterias to lay out an office floor. That is 10x to 100x leverage for the architect, plus automation of the rote path from concept to something you can walk into. There is still no map of every building on the internet, so none of that sits fully inside frontier models yet.

Labenz wonders whether architecture is a profession where "we will just want more" fails. Graphic design is ephemeral; buildings are bottlenecked by construction cost. If architects become 100x more efficient, we will not build 100x more buildings, and homeowners might do concept-level work themselves. Should architects worry? Milinovich thinks not. Something built in the real world still has to be signed off by an accredited architect who is legally liable. The more interesting pressure is the bid. On massive commercial projects, and often on residential ones, a huge share of time is spent doing the work before the client chooses you. AI will first eat that permutation space: visualize the client's idea with the architect's stamp on it, and win the deal. After a finished BIM model, permit packets and electrical and plumbing drawings are discrete automation. The creative-director role gets heightened, not deleted. Liability and insurance may eventually let a software firm take that risk; the near-term version is semi-prefab housing, a design system of frames, floors, and roofs, with architecture and construction vertically integrated. Aesthetic started as that idea before moving into graphic design. Katerra tried a version and failed. He does not think the lesson is that the idea is bad; it is execution and AI enablement. Anything that brings down the cost of housing, Labenz adds, would be welcome.

Milinovich's blog post, "The GenAI Application Layer," is the last third of the argument. Every major platform shift, on-prem to cloud, desktop to mobile, pre-AI to post-AI, aggregates value first at infrastructure, then at platform, then at application. Cloud made that visible: chips and data centers, then hyperscalers, then almost every internet company. He thinks AI is still in the infrastructure era. Global GDP is on the order of $100 trillion, headed toward about $130 trillion by 2030; tech is about $5 trillion of that. If tech's share of the service-heavy remainder grows even three percentage points, that is a multi-trillion-dollar application-layer opportunity. Sequoia's essay on a $600 billion AI gap, in his reading, is evidence the application layer is underserved relative to the labor that will be augmented or automated.

The post names GitHub Copilot, Jasper, and Harvey. Labenz is not equally confident in all three. Jasper became the poster child for a thin wrapper that found a customer and got to market, then faced labs shipping a general canvas OpenAI will not need ten years to copy. Harvey is a hard vertical, and GPT-5 could still scramble it. Canva, by contrast, has a decade of design software a lab will not recreate next week. What leading indicators does he want? First, deep customer centricity: a specific person whose eyes you can see and whose pain you can feel. As the platform layer builds its own thin, generalized apps, opinionated vertical focus matters more than ever. Second, go where tech has not already eaten the world: mining, real estate, construction, places with immediate leverage. Third, design UX that knows the jagged edge. Off the straight-and-narrow, AI will give meaningful value and not a 100 percent solve; the product has to cover the shortfall. Those three, he tells founders, are what make him think a team is set up for the next decade.