Beyond the First Image: Canva’s Case for Editable, Iterative AI Design

1. Guest Background

Stefano Caraza, Canva’s head of AI research, joins host Sam Charrington on The TWIML AI Podcast for a roughly 57-minute discussion about making AI useful throughout a design task. The episode examines layout control, revision, visual agents, and the cost of running repeated model calls. Its slot-machine framing describes the uncertainty of obtaining and then changing a generated result; it is the problem the conversation investigates, rather than a measured assessment of every image model.

Caraza speaks from the research and product work he describes at Canva, a company providing design tools and integrating AI into creative workflows. His team’s relevant work includes Craft, an image-generation model, and Brush, a vision-language model for creating and editing designs. He says Canva had approximately 100 AI models in production at the time, including smaller systems. That context matters because his argument concerns a collection of capabilities embedded in an editor, rather than an isolated image-generation demonstration.

The two featured models also illustrate different development choices. According to Caraza, Craft was trained from scratch, while Brush builds on an open-weights model. The interview therefore supplies experience with both proprietary training and adaptation of an existing foundation. The supplied evidence establishes his role and these areas of work, but does not provide a broader career history. His account is best read as a research leader’s explanation of his team’s decisions and observations, with the supporting detail and uncertainty assessed accordingly.

2. What the Episode Covers

Craft separates the path from a request to an image into three stages: enriching the input and context, producing a structured scene description, and generating pixels. Caraza says the components are trained individually and also examined together. The intermediate representation remains text-based, but includes bounding boxes and relative and absolute positions. Layout consequently becomes something that an agent or a person can inspect and change before rendering. His example is to propose ten arrangements of a page, obtain human confirmation, and only then generate the images.

Editing existing material provides a second route into that workflow. Craft allows people to resize or reposition elements before regenerating pixels. Magic Layers decomposes a flat image into layers and elements; Caraza says most images processed with it originate outside Canva. His example of asking an image model to move something two feet captures a practical ambiguity: the image may contain no reliable basis for that distance. Moving an element directly gives the designer a more explicit way to express the intended change.

The training account combines data selection with coordination between components. Caraza describes permissioned user data, externally licensed material, stock imagery, synthetic data, and millions of designs Canva created or licensed. Diversity is an explicit goal, both to reduce repetitive-looking outputs and to serve different markets and languages. He also says scene descriptions became substantially more detailed over six months, while the diffusion model grew somewhat to absorb that information. The planning system and renderer had to be calibrated together; adding detail upstream was only useful if the downstream model could realize it.

Brush operates at the level of a design document containing text, vector graphics, images, and multiple layers. It can use design tools to rearrange a composition, improve readability, and respond to abstract requests about emphasis or style. Caraza’s larger example is a cohesive, branded thirty-slide deck. This shifts the discussion from generating a single image to choosing operations across heterogeneous elements and maintaining relationships within a document.

Access beyond Canva’s interface is another part of the episode. Caraza says Craft was available through Canva Connect API and Canva MCP, with further variants planned for MCP. Existing exposed capabilities included Magic Layers and Lucid Origin. He identifies enterprise-specific workflows and agent-driven use as the principal sources of demand, with generated work able to return to Canva for editing. He also reports rapidly growing MCP usage and triple-digit year-over-year enterprise growth. Without detailed statistics in the interview, those statements describe his view of demand; they do not establish adoption volumes, customer outcomes, or returns from any particular integration.

3. Core Views: Reasoning, Examples, and Limits

The central argument is that design AI should be organized around revision. Caraza estimates that AI gets a task roughly 70–80 percent of the way there and that a person remains necessary for the final stretch. He contrasts minute-long generation waits with the responsiveness he wants in everyday creation, invoking ten seconds as a game-development reference. These are experiential judgments, not measured completion rates or universal patience thresholds. Their analytical value lies in the dependency they expose: when intent must be clarified through repeated changes, each delay becomes part of the total task. An impressive first result cannot, by itself, establish a usable creative process.

Structured planning gives that argument an operational form. Comparing ten layouts before rendering lets someone judge information arrangement without first committing to finished pixels. Directly moving an element addresses the ambiguity in Caraza’s two-feet example. The broader reasoning is that a useful representation preserves decisions that people may want to revisit. Yet structure alone does not guarantee fidelity. His account of expanding scene detail while increasing diffusion capacity makes the condition explicit: the renderer must be able to implement what the planner specifies. A richer description is not automatically a better image.

The case for specialized models is similarly conditional. Caraza attributes Canva’s opportunity to distinctive data and deep integration with its product stack. He reports that task-specific models can outperform larger alternatives with two orders of magnitude fewer parameters, and puts Canva’s comparable design-task cost at about one-fortieth of the most expensive frontier models. The interview supplies no reproducible comparison identifying models, task sets, or accounting assumptions. His judgment that a properly tuned planning LLM or VLM of around 100 billion parameters or less can handle most creative tasks also has a stated possible exception in video understanding. The approximately one-dollar-per-image figure is an illustration of economic pressure, not a published price measurement. Together, these claims motivate testing specialization; they do not establish a universal optimum for model size.

Low latency could also change how quality is produced. Brush can inspect what it generates and modify it again. Caraza suggests that a sufficiently fast loop might perform about ten iterations before presenting work to a user. This is more consequential than making a progress indicator disappear sooner: it expands the strategies an agent can afford to attempt. But the episode provides no measured quality gain from ten passes, and additional cycles need not converge on the right result. An agent with an inadequate criterion could repeatedly optimize the wrong feature while consuming the available time and inference budget.

That makes evaluation a central constraint. Caraza agrees with Charrington’s reference to Craft ranking around fourth on Design Arena at the time, while arguing that public benchmarks emphasize initial generation. His example, Present Bench, contains about 200 presentation-related tasks by his account, but still primarily evaluates new outputs. Correctly changing an existing design, preserving its coherence, and doing so quickly and cheaply are different requirements. Canva uses articulated rubrics for readability, layout, and composition with VLM judges, alongside human evaluation. The important dependency is that people must first express what the judge should assess. A score cannot resolve an undefined design goal.

The cartoon example further limits what a headline ranking can explain. Caraza reports stronger performance on portraits, people, and natural landscapes, but says Craft was outside the top three for cartoons. He suspects that work emphasizing photorealism made early cartoon outputs overly realistic and three-dimensional. Because the team believes relevant content already existed in pretraining, it is adjusting style through fine-tuning and reinforcement learning. The observed category difference and the proposed explanation should remain separate: the interview does not independently demonstrate that photorealism optimization caused the weakness.

Finally, Brush’s tooling experience argues for matching an interface to the model’s existing competence. Caraza describes a third-generation harness becoming thinner and simpler, allowing code-driven tool use with visibility into results. Moving, rotating, and enlarging an object can be expressed as a complete state, a diff, or editing API calls. The team considers which forms fit the model’s training distribution. The transferable reasoning concerns reliable expression and feedback, rather than an assumption that a novel design language or a minimal harness will always be superior.

4. Learning and Application

A design-product team can apply the episode by changing its evaluation unit from a generated image to a completed revision. Start with an existing page and request an element-size change, a readability improvement, or consistent styling across several pages. Check whether the requested change occurred, whether unrelated content was disturbed, and what the entire task cost in time and inference. Following Caraza’s rubric approach, score readability, layout, and composition separately and retain human review. A visual feedback loop then becomes a testable intervention. His ten-iteration example should be treated as a candidate strategy, with a stopping condition and a budget, rather than a required number of passes.

For agent interfaces, compare complete-state output, local diffs, and editing API calls on the same move, rotate, and resize tasks. Examine whether the model expresses the operation correctly, whether the tool executes it, and whether the returned result supports correction. This is an application of the interview’s reasoning, not a reported experiment with a declared winner. A simpler interface may be sufficient when it already supports dependable actions. Conversely, simplifying the harness should not remove the feedback that lets an agent determine what happened.

Inference work should begin with a bottleneck assessment and an inventory of what the team controls. Caraza connects efficiency to AI access for free users, including students and nonprofits. His list spans hardware, kernels, model steps, distillation, serving, batching, and caching. Canva, he says, runs and optimizes models on cloud infrastructure without owning the physical hardware. Deeper changes depend on access to weights, self-hosting, and collaboration across AI research, graphics engineering, and cloud scaling. Teams without those capabilities can still compare task costs and controllable service choices, but cannot assume Canva’s full-stack optimization opportunity is available to them.

A separate exploration budget can make speed-versus-fidelity choices explicit. At the time of the interview, Craft Fast and Craft Ultrafast had been announced as forthcoming variants; Caraza positions faster options for prototyping and the highest-fidelity model for work requiring greater quality. Because Craft contains planning and rendering components, distillation should be evaluated across their interface, not only by timing either component alone. His report that generating a second of 720p video fell from roughly half a dollar to below one cent lacks comparable-quality, hardware, and accounting details. The 0.1-cent figure is a prospective adoption threshold in his argument. Neither figure is a sufficient basis for a procurement commitment or capacity forecast without local measurement.

For services used by agents, review the entire path from finding documentation to understanding an interface, trying it, validating a prototype, and requesting expanded use. Caraza favors structured information and trials that do not require subscribing first, followed by human approval before scaling. Lower trial friction and authorization for wider usage are distinct product decisions. Research planning can follow the same separation between stable purpose and changeable implementation. His team keeps accessibility, iterative creation, and workflow obstacles in view while reprioritizing projects as opportunities change. His statement that the models being released had not been underway a year earlier illustrates that team’s adaptation; it does not prescribe the same project cadence for every organization.

Source

• Original episode: AI Design Still Feels Like a Slot Machine