Miju Labs

← All buyers · Canva (incl. Leonardo.ai)

D

Danny Wu

Head of AI Products, Canva · Canva (incl. Leonardo.ai)

medium confidenceTier A · Talk to firstImage, video & audio modelsUI · image

Owns Canva's AI product surface — Canva AI, Magic Studio and the Canva AI 2.0 release built on Canva's own design foundation model. He is the person who publicly described what that model is trained on: millions of professionally created templates already separated into editable components. He decides what counts as a good generated design in product terms, and is the natural requester for design-quality judgement data.

Background

Roughly a decade at Canva, moving into the head-of-AI-products role as Canva industrialised AI across the editor. His framing is unusual and useful: Canva does not want to generate images, it wants to generate designs — layered, editable, structurally correct — which means its quality bar is layout, hierarchy and editability rather than photorealism.

He also holds Canva's model-agnostic line: Canva ships partner models from OpenAI, Anthropic, Google and Black Forest Labs alongside its own. That matters commercially — Canva buys data to differentiate the design layer, not to win a pixel-quality race.

What they run now

Canva AI 2.0 (Oct 2025) and the design foundation model underneath it; AI features across the editor; partner-model integration. Canva's Generative Experience Group builds the design-generation models, and a Head of AI Research role posted to bridge Canva and Leonardo.ai research reports to the CTO and the Gen AI lead.

What they buy

Inference, with good supporting evidence: judgements about design quality — layout, hierarchy, alignment, readability, brand consistency — as preference pairs and as rubric-scored critiques; template and layered-design structure data; and evaluation sets that distinguish 'a correct design' from 'a pretty image'. Canva's own supply is its template library and a 2023 $200M creator fund that pays designers who consent to training use, so the gap a vendor fills is judgement and evaluation, not asset volume. No external vendor named publicly; Canva declined to say whether customer data trains the design model.

Career

  1. presentHead of AI Products, Canva — ~10 years at Canva per interview

On the record

what the design model is trained on · 2026-01

Says Canva trained on millions of professionally created templates already separated into editable components, so the system generates designs rather than images. source

human control in the loop · 2026-01

Argues the AI parts and human control and creativity should mesh and work in harmony instead of running on two separate tracks, so users do not have to keep re-prompting. source

speed vs quality · 2026-01

Says Canva is focused on making its AI as fast as possible without trading off quality. source

model agnosticism · 2026-01

Describes Canva as model-agnostic, partnering with OpenAI, Anthropic, Google and Black Forest Labs while investing in internal models. source

Interviews & talks

Why it matters here

Canva is one of very few buyers whose model target is design correctness rather than image aesthetics, which is precisely Julian's niche. Wu is the product owner of that target and the person who articulates it publicly, so he is the person to whom a 'we measure design quality' pitch is legible. He is also likely to route a serious conversation to the Generative Experience Group or the Head of AI Research.

How to reach

Speculation: open with the evaluation gap implied by his own framing — if the model generates designs, not images, then image-preference benchmarks cannot score it, and there is no public benchmark for layered design correctness. Offer a small designer-panel evaluation of Canva AI 2.0 outputs against a layout/hierarchy rubric as a free artefact. He is active on podcasts and writes on the Canva newsroom; a substantive public post on design evaluation is a plausible way to get on his radar.

What we could not establish
  • No X handle verified.
  • Whether Canva uses any external annotation or preference-data vendor today.
  • Whether the $200M creator fund covers training-data consent for the design model specifically.

Why they are on the map

Owns Canva AI; described design-model training on professional templates. source — evidence dated 2026-01

Needs the data for a model; writes the spec.

How Canva (incl. Leonardo.ai) buys data

Design model trained on 'millions of professionally created templates' (Danny Wu, Jan 2026); $200M creator fund (2023) pays royalties to designers who consent to training use. Canva declined to say whether customer data trains the design model (Computerworld, Oct 2025). Hiring Head of AI Research to bridge Canva and Leonardo research; 'Generative Experience Group' runs design-generation ML.

AI products under Danny Wu (Head of AI Products); Generative Experience Group builds design-generation models; Head of AI Research reports to CTO Brendan Humphreys and GenAI lead Paul Thompson and works with Leonardo.ai's research team in Sydney.

Known vendors: Canva Creators (template royalties)

Sources

  1. The AI Innovator interview with Danny Wu (Jan 2026)
  2. Computerworld: Canva debuts foundational 'design' model (Oct 2025)
  3. Canva Newsroom author page