Leonardo.Ai wants to fix AI image generation's control problem

Anyone who uses AI image generators knows the slot machine fatigue: you type a paragraph of description, hit generate, and pray the model doesn't ignore half your nouns.

CNET reported that creative platform Leonardo.Ai is testing Phoenix, its first entirely in-house foundation model. Fresh off announcing its acquisition by design giant Canva on Tuesday, the startup is betting big on turning prompt roulette into predictable design work.

Founded in October 2022, Leonardo.Ai has quietly amassed 17 million registered users and saw 4 million to 5 million monthly active users in May. It also raised $31 million last December from backers like Samsung Next and Blackbird.

Ditching prompt roulette

Until now, Leonardo's tools relied on open-source Stable Diffusion backbones trained on public domain photos, Creative Commons media, and synthetic data.

Phoenix marks a deliberate shift. Scheduled for a broader launch later in the month, the in-house model is engineered for extreme prompt adherence. According to Leonardo, the model handles long, highly descriptive queries without dropping specific details along the way.

That foundation joins a toolbox already laser-focused on granular control. Leonardo previously launched Realtime Canvas for brush-stroke rendering, Realtime Gen for typing-speed generation, and Edit With AI for local modifications. The platform also lets users train custom micro-models using their own art, with over 470,000 community models built as of December.

While a basic tier is free, Leonardo's core workflow hooks—like character reference, style reference, and custom model training—are gated behind paid plans starting at $10 a month.

In hands-on testing, CNET reviewer Katelyn Chedraoui noted that Leonardo follows instructions noticeably better than rivals like Midjourney, though it still lacks deep post-generation editing features and occasionally misses that elusive "wow" factor.

Why it matters

We think the AI image space is finally shifting from party tricks to production pipelines.

If you are a concept artist or graphic designer, Midjourney's cinematic accidents don't pay the bills when an art director asks for the exact same character in three different angles. Leonardo's relentless focus on prompt compliance, reference locks, and custom training targets precisely the friction points that keep pros from relying on AI daily.

Strategically, building Phoenix and selling to Canva solves Leonardo's biggest existential threat. Wrapper startups building UI around base Stable Diffusion are vulnerable to commoditization. By controlling its own weights and plugging directly into Canva's millions of enterprise and casual creators, Leonardo gets distribution that standalone rivals can only dream of. The challenge now is matching Adobe's seamless suite-level editing without alienating creators who just want a single, stunning image without tweaking ten control sliders.