The TWIML AI Podcast

Stefano Corazza on why creative AI is a loop, not a one-shot

Stefano Corazza· Head of AI Research at Canva
·~57 min·English·TWIML
MultimodalInferenceTrainingAgentsAI Company
TL;DR

Canva's head of AI research argues that useful creative AI comes from a fast, controllable loop rather than one-shot generation: own your data and train smaller models, separate layout from pixels, judge with rubric-driven VLMs plus human review, and keep driving inference cost down because the field is far from its plateau.

01Core stance

AI gets you 80%, not 100%

The last mile belongs to a human, so the whole product has to be built around fast iteration — not a single perfect generation.

You have to build around iteration. AI will get 70% 80% there. I will never get to 100%. And so, you have to factor in that for the last mile, you need a human in the loop and be able to iterate in a way that is fast and also fun.

— Stefano Corazza, The TWIML AI Podcast
Key Insight
Treating the human as a permanent part of the pipeline, not a fallback for when the model fails, reframes the whole engineering problem: the goal shifts from a better one-shot output to a faster, cheaper loop.
02Build vs. buy

A model 100x smaller, 40x cheaper

Unique data plus deep vertical integration let a smaller, self-trained model beat a frontier one on the same task — at a fraction of the cost.

right now we are in the you know 40x cheaper than like the most expensive frontier models that can do the same design task but they are using you know orders of magnitude more parameters to do it

— Stefano Corazza, The TWIML AI Podcast
Key Insight
The argument generalizes beyond Canva: any company sitting on proprietary, domain-specific data and willing to own its inference stack can likely out-economize a general frontier model on its narrow task.
03Architecture

Design the layout, then paint the pixels

Most of the model's work is building a structured scene graph; turning it into pixels is only the last step — so edits happen before anything is baked in.

we have created like an image model that most of the work actually is creating a scene graph of what the design should look like. And then the conversion to pixel is just the last step that we train from scratch.

— Stefano Corazza, The TWIML AI Podcast
Key Insight
Separating structure from rendering is what makes the 'move this two inches' request answerable: a flat image has no notion of an element, but a scene graph does, so a human or an LLM can nudge a layout without re-rolling the whole picture.
04Scaling

Creative work hits a ceiling near 100B

For creative tasks, the input model stops improving well before frontier scale — roughly 100 billion parameters is already enough.

you don't need more than 100 billion parameter to do most of the creative tasks I would say you know probably like video understanding maybe needs just a little bit more because it's probably the most complex type of task

— Stefano Corazza, The TWIML AI Podcast
Key Insight
If quality plateaus at a size the team can afford to run itself, the race shifts off raw parameter count and onto data quality, post-training, and inference cost, which is exactly where an integrated company has the edge.
05Evaluation

If you can't articulate it, the VLM can't score it

A vision-language model can judge design quality, but only against a rubric humans were able to write down first.

it forced us also to really define what a good design means right and so if humans are not able to articulate that then VLMs are not going to be able to score it

— Stefano Corazza, The TWIML AI Podcast
Key Insight
Building the evaluator forced a harder question than building the model: the team had to make 'good design' explicit enough to score, which turns a matter of taste into measurable rubrics like readability and layout.
06Agent design

A thinner harness, a freer model

Brush is an agent, and its hardest part was the harness — made thinner each generation so the model can write code, use tools, and iterate on its own output.

the high level goal is basically to give as much freedom to the LLM to really run free of constraint and really be able to like tap into the tools

— Stefano Corazza, The TWIML AI Podcast
Key Insight
The design heuristic is to 'align with the wind' — express operations in the formats the model already saw most in training, rather than inventing a bespoke language, so the harness leverages the model's existing grooves instead of fighting them. A faster loop then lets the agent run many iterations before a human ever sees the result.
07Inference economics

Far from the plateau

Squeezing a 10x inference gain in a single month is the sign you are still early — and owning the full stack is what makes those gains possible.

if you can squeeze 10x performance improvement in a month means you're very early right like we are so far from the from the plateau right now

— Stefano Corazza, The TWIML AI Podcast
Key Insight
Framing cheap inference as a reliability and access goal, not just a margin goal, connects it back to the loop: lower cost per generation is what lets free users and self-iterating agents run many attempts instead of rationing them.
08The next user

Building for agents, not just humans

Agents need structured information, easy previews, and low-friction access — so Canva is building for them alongside its human users, as demand grows for API and MCP calls.

we are basically building not only for humans but also for agents

— Stefano Corazza, The TWIML AI Podcast
Key Insight
Designing for an agent means designing for a caller that reads documentation instead of a UI and decides on its own which tool to use, so structure, previewable trials, and low sign-up friction become product features rather than developer niceties.