Víctor Perez

Which Víctor Perez is this?

The name Víctor Pérez is common in the Spanish-speaking world, and several public figures share it. Before anything else, this profile establishes exactly who is being documented — and who is not.

Quick overview

Who is Víctor Perez?

Víctor Perez is a Spanish machine learning engineer who became one of the most visible product minds in generative AI without ever finishing a graduate degree. He is the co-founder and CEO of Krea, a San Francisco company that builds what it calls a creative interface: a browser workspace where images and video are generated, steered and refined in real time, across many AI models at once, rather than one prompt at a time.

His path is unusual for an AI founder. He came into machine learning sideways — from music, graffiti and audiovisual engineering rather than a computer science PhD track — and that origin shows in everything Krea has built. Where much of the industry optimised for benchmark scores, Perez consistently optimised for the distance between having an idea and seeing it on screen.

By 2025, that bet had produced a company valued at $500 million with more than 20 million users. By 2026, Krea had gone further still, moving from a layer on top of other people's models to training foundation models of its own.

Early life and childhood

Perez was born and raised in Barcelona. The most consistently documented fact about his youth is that he was, first and foremost, a musician. From roughly the age of twelve until his late teens he played in a band, and by his own account in a Universitat Pompeu Fabra interview, everything he did was creative — he composed the music and produced it himself. Alongside music he drew, illustrated and worked with graffiti.

That combination — sound, image and the machinery underneath both — set up the decision that shaped his career. At eighteen, as Forbes reported, he faced a fork: enrol at a Spanish conservatory to study classical guitar, or go to university and study technology. He chose neither exclusively. He looked for a degree that contained both.

What actually influenced him

• Music production — an early, practical apprenticeship in creative tools, and in the frustration of learning them.
• Graffiti and illustration — a visual vocabulary his later co-founder would recognise; Bain Capital Ventures noted that Perez's love of graffiti art and graphic design was part of what deepened his friendship with Diego Rodriguez.
• Mathematics and physics — Perez told TechCrunch he was drawn to "physics and maths and problems that challenged my mind" as a student.

Education

Universitat Pompeu Fabra — Barcelona, Spain

Institution → Universitat Pompeu Fabra (UPF) · City → Barcelona · Country → Spain · Degree → Bachelor's in Audiovisual Systems Engineering, Poblenou campus.

He chose the programme precisely because it sat between music and mathematics. In his UPF interview he described it as teaching how a microphone works internally, how audio codecs function, how an image is encoded, and how sound behaves physically and mathematically.

Programming, not audio, became the hook. He described finding programming intensely creative — one tool that could be pointed at generative art in Processing or at any application he could imagine — and he built deliberately useless practice apps as a form of expression while learning.

The professors who turned him toward AI

Perez has publicly credited two people at UPF — Coloma Ballester and Adrià Arbués — with drawing him into deep learning and machine learning. The moment that stuck, by his own description, was realising a computer could understand the structure of images well enough to produce new ones that had never existed.

He also met Diego Rodriguez on this degree, around 2015. Perez credits Rodriguez with sparking his interest in AI, including his introduction to generative adversarial networks.

Recurse Center — New York City, USA

After graduating, Perez won a place at the Recurse Center, a highly selective, self-directed programming retreat in New York. It ran roughly three months. He described the experience as monastic: like being a Buddhist monk, except that instead of meditating he was studying artificial intelligence. UPF's festival profile lists him as a Recurse Center alumnus.

Cornell University — Ithaca / New York, USA

In 2022 Perez received a full graduate fellowship to study at Cornell. He attended for approximately one day before abandoning it to build Krea.

Perez is frequently described as a "Cornell alumnus" on aggregator profiles such as The Org and LinkedIn summaries. In practice, he did not complete a Cornell degree — the fellowship was abandoned almost immediately. This is one of the most commonly repeated inaccuracies about him.

Early career: six years inside generative models

Between graduating and founding Krea, Perez spent roughly six years working on generative models — long before the phrase "generative AI" entered common use. The sequence matters, because each step supplied a different piece of what Krea would eventually become.

The idea that ran underneath all of it

Across interviews, one thesis repeats. Perez has described what excites him about the technology as the possibility of minimising the friction between an idea and its real-world representation. Professional creative software, in his framing, forces people to spend their attention on how to do something instead of what to do. Every project he touched — a painting interface, a sculpting tool, a design canvas — was an attack on that same gap.

What Víctor Perez actually built

Perez is not the inventor of diffusion models, CLIP, or text-to-image generation. Attributing those to him would be false. What he did was invent, popularise and productise a specific way of using them: continuous, real-time, human-steered generation — and later, the models purpose-built to make that interaction possible.

Creation 1 — ProsePainter (2021)

Creation 2 — "Sculpting with Words" (NeurIPS 2021)

Creation 3 — Krea (2022 – present)

How Krea was actually created — step by step

1. The fork (2022). Perez flew to the United States to begin his Cornell fellowship. He left after a single day, convinced the moment to build was now, not after a degree.
2. The pitch. On arriving, he pitched Rodriguez on a browser-based generative art platform. Rodriguez left his own studies to join.
3. The application neither of them planned. Rodriguez had already applied both of them to HF0 — an intensive residency described as a Y Combinator for technical founders — using Perez's concept, without telling him.
4. The residency. Despite an interview the pair themselves rated as poor, HF0 accepted them, offering a residency place alongside roughly $500,000 for about 3% equity.
5. Incorporation. Krea was formally incorporated in March 2022, during the residency.
6. Build on other people's models first. The early product ran on modified Stable Diffusion. Perez was explicit about the reasoning: they did not want to train models from scratch until they knew exactly what people needed.
7. Find the real product inside the demo. The breakthrough was latency. Once generation became fast enough to feel continuous, the product stopped being a generator and became an instrument.
8. Then build the models. Only after product-market fit did Krea start training its own: Krea 1, then Krea Realtime 14B, then Krea 2.

The technology, explained at three levels

Level 1 — For anyone

Most AI image tools work like ordering food: you write a description, wait, and receive whatever arrives. Krea works like drawing next to someone who is drawing along with you. As you move your mouse or change a word, the picture updates almost instantly. You are not describing an outcome and hoping — you are steering one.

Perez's own analogy from his UPF interview is the cleanest version of this: computers let us digitalise the paintbrush so we could paint with pixels; AI lets us build tools where you interact with a painter instead. How you talk to the painter matters more than how you hold the brush.

Level 2 — For practitioners

• Multi-model orchestration. Krea aggregates more than twenty third-party models — including Black Forest Labs' FLUX family, Ideogram, Google's Gemini image models, Recraft and the Seedream series — behind one interface, alongside its own. The routing layer, not the model, is the product.
• Real-time generation. The live canvas is driven by heavily distilled models. Where conventional diffusion pipelines run fifty or more denoising steps, Krea's real-time path uses Latent Consistency Model techniques requiring only two to four steps, targeting sub-50-millisecond updates.
• Customisation via LoRA. Enterprise users train lightweight adapters on their own visual identity. Adapter files typically run 50–200MB and steer the base model toward a trained style, character or object. Krea's in-app training entered beta on 21 May 2026 and opened to all users on 1 June 2026.
• Distribution as strategy. Releasing open weights on Hugging Face gave Krea research credibility and an ecosystem it did not have to build.

Level 3 — Model-level detail

Perez is a listed author on both the FLUX.1 Krea [dev] release and the Krea 2 technical report — notable for a CEO of a company at this stage, and consistent with his history as a hands-on researcher rather than a purely commercial founder.

The hard parts

The professional record

Companies and organisations

Research papers and technical publications

Patents

No reliable public source was found to verify any granted patent or published patent application naming Víctor Perez of Krea as inventor. His documented contributions are published research, open-weight model releases and shipped product — not patent filings. Any claim of a specific Perez patent should be treated as unverified until a patent number is produced.

Key collaborators

• Diego Rodriguez Prado — co-founder and CTO. Met Perez at UPF around 2015; introduced him to coding practice and generative adversarial networks; applied both of them to HF0 without telling him.
• Coloma Ballester and Adrià Arbués — UPF academics Perez credits with pulling him into deep learning.
• Black Forest Labs — foundation-model partner on Krea 1 and FLUX.1 Krea [dev]; provided the base weights on which Krea's post-training ran.
• Krea's research team — including Sangwu Lee, Erwann Millon, Le Zhuo, Titus Ebbecke, Will Beddow, Gabriel Menezes, Mihai Petrescu and Gian Saß, co-authors with Perez on the FLUX.1 Krea [dev] and Krea 2 releases.
• Investors — Bain Capital Ventures (led the Series B, partner Aaref Hilaly), Andreessen Horowitz, Abstract Ventures, HF0, A*, Gradient Ventures, plus angels including Keith Adams and Pamela Vagata.

Awards and recognition

• Forbes 30 Under 30 — Media (2026 list), published December 2025.
• "la Caixa" Foundation graduate fellowship (2022), with recognition presented by King Felipe VI.
• Recurse Center selection.
• Forbes Next Billion-Dollar Startups — Krea listed (2025).
• Speaking: +RAIN Film Festival (UPF, June 2023), Europe's first AI-generated film festival; This Week in Startups podcast (episode E1850).

What he is doing now (as of mid-2026)

Perez remains co-founder and CEO of Krea in San Francisco, leading a team that had grown to roughly 103 people by June 2026. The strategic shift of the past year is the significant one: Krea stopped being only an interface company.

• March 2026 — a full application redesign: unified navigation, transparent model selection, drag-and-drop workflows, a customisable workspace, a rebuilt mobile experience, and an iPad voice mode that updates the real-time canvas as the user speaks.
• May 2026 — the Krea 2 foundation model family begins rolling out; in-app LoRA training enters closed beta for Max and Business subscribers on 21 May, opening to all users on 1 June. A Krea 2 API arrives with fal.ai as a day-zero partner.
• June 2026 — Krea 2 Raw and Turbo released as open weights under a custom licence, accompanied by a detailed technical report on which Perez is a listed author.
• Enterprise focus — clients including Shopify and Perplexity now fine-tune private models on their own aesthetics.

The idea he has been chasing since he was a student

There is a single sentence underneath everything Perez has built, and he has been saying versions of it for years: the point of this technology is to shrink the distance between an idea and its representation.

It comes from somewhere specific. A teenager producing his own music learns that the hardest part of making something is rarely the idea — it is the six months spent learning which knob does what. Professional creative software, as he has described it, forces you to think about how to do something when you should be thinking about what to do.

Watch how that thesis produces each artefact in turn. ProsePainter: if a prompt box is the wrong interface, let people paint the prompt. Sculpting with Words: if that works in two dimensions, put the user's camera inside the optimisation loop in three. Krea: if generation is fast enough to feel continuous, the interface stops being a form and becomes an instrument. Krea 1 and Krea 2: if off-the-shelf models produce a generic look, train models whose defaults are opinions.

His view of graphic design history explains the ambition. Each revolution — paper to pixel — was fundamentally about how quickly an idea becomes something visible. He has argued that AI drives that friction toward zero, which forces creative software itself to change shape. His stated goal is an interface with context about who is using it and why, that can reason and act on that basis.

On ethics, his position is consistent and not especially convenient for his own industry: regulation should focus on how people use the technology rather than on the technology itself, because norms are learned through trial and error rather than decided in advance around a table. He has been explicit that style imitation is a real copyright infraction and that indistinguishability from real images is a coming problem.

Verified quotes

"we're making technology more human"— Víctor Pérez, interview with Universitat Pompeu Fabra, 25 May 2023, on how AI changes the relationship between people and computers.

"I liked physics and maths and problems that challenged my mind"— Víctor Perez to TechCrunch, 7 April 2025, recalling his student days in Barcelona.

"enabling people who didn't think of themselves as creatives"— Víctor Perez to Forbes, on who generative AI actually empowers.

"It's always made by you with AI, never just made by AI."— Víctor Perez, quoted by Bain Capital Ventures, April 2025, describing Krea's mission.

Timeline

Why the work matters

What he is best known for

For turning generative AI from a slot machine into an instrument — and for proving that the interface layer above the models can be a defensible business, not a thin wrapper.

25 verified facts

1. He was born and raised in Barcelona.
2. He played in a band from roughly age twelve to his late teens, composing and producing the music himself.
3. At eighteen he nearly chose a classical guitar conservatory over university.
4. Graffiti and illustration were formative creative outlets before engineering was.
5. He picked Audiovisual Systems Engineering at UPF specifically because it merged music production with mathematics.
6. He describes programming — not audio engineering — as what hooked him during the degree.
7. He deliberately built non-functional practice applications as a form of self-expression while learning to code.
8. He credits UPF academics Coloma Ballester and Adrià Arbués with drawing him into deep learning.
9. He met co-founder Diego Rodriguez at UPF around 2015 — years before the generative AI boom.
10. Rodriguez, not a professor, introduced him to generative adversarial networks.
11. He attended the Recurse Center in New York on a scholarship for self-directed programmers.
12. He compared that three-month retreat to being a Buddhist monk — studying AI instead of meditating.
13. He led machine learning research at Plyzer Intelligence in Barcelona from 2018 to 2020.
14. That role was about graph neural networks and knowledge graphs for e-commerce — not image generation.
15. He has collaborated on creative technology projects with the NYC Media Lab and with CHANEL.
16. He built ProsePainter, a tool that let users paint text prompts directly onto a canvas.
17. His NeurIPS 2021 paper "Sculpting with Words" used CLIP and a differentiable renderer to sculpt 3D geometry from text.
18. In that system, the user's own camera position determined which viewpoint the optimisation targeted.
19. He won a "la Caixa" Foundation graduate fellowship in 2022, recognised by King Felipe VI.
20. He attended Cornell for approximately one day before leaving to found Krea.
21. Rodriguez applied both of them to the HF0 residency using Perez's idea — without telling him.
22. They rated their own HF0 interview as poor. They were accepted anyway.
23. Krea deliberately avoided training its own models for years, on the reasoning that they should not train from scratch until they knew what users needed.
24. He is a listed author on both the FLUX.1 Krea [dev] release and the Krea 2 technical report, despite being CEO.
25. Krea reached more than 20 million users before its Series B was publicly announced.

Myth vs fact

What his career actually teaches

1. Creative background is a technical advantage. His instinct for the interface came from years of struggling with music software, not from a research lab.
2. Solve the friction, not the benchmark. Everything he built attacked the gap between idea and image; none of it started from a leaderboard.
3. Don't build the expensive layer first. Krea rented models until it knew what to train, and only then spent on foundation models.
4. Latency is a feature. The single decision that defined Krea's identity was making generation fast enough to feel continuous.
5. Credentials are optional; conviction is not. A one-day Cornell career is the sharpest illustration of the trade he made.
6. Partner where you can't outspend. Collaborating with Black Forest Labs bought Krea a frontier base model it could never have trained alone.

Frequently asked questions

Sources

Tier 1 — Primary and official

• Universitat Pompeu Fabra — interview with Víctor Pérez, Faculty of Communication, 25 May 2023
• Universitat Pompeu Fabra — +RAIN Film Festival participant profile
• Krea — Krea 2 Technical Report, 23 June 2026
• Krea — FLUX.1 Krea [dev] official repository
• Black Forest Labs — FLUX.1 Krea [dev] release announcement and Hugging Face model card
• Morphogen Research — "Sculpting with Words", NeurIPS 2021
• Bain Capital Ventures — investment note on Krea, April 2025

Tier 2 — Established press

• TechCrunch — Ingrid Lunden, "Krea raises $83M to be the one-stop shop for GenAI creatives", 7 April 2025
• Forbes — 30 Under 30 Media 2026 and Krea company profile
• VentureBeat — Krea 2 Raw and Turbo open-weights release, June 2026

Tier 3 — Databases and aggregators (used only for corroboration)

• Crunchbase, PitchBook, Tracxn, The Org, LinkedIn public profile

The distance between an idea and an image

Víctor Perez's career makes sense only if you read it backwards from a single obsession. A teenager in Barcelona learns that the hardest part of making music is not having the idea — it is the software between the idea and the sound. Fifteen years later he is training foundation models whose entire purpose is to delete that distance.

Everything in between is the same problem restated at increasing scale: a painting brush that carries language, a 3D mesh that reshapes under a sentence, a canvas that redraws faster than you can second-guess it, and finally models trained to have opinions so that ordinary users inherit taste they never had to learn.

He did not invent generative AI. He did something more specific and, arguably, more useful — he kept insisting that the person should stay inside the loop. In an industry that spent years asking how good the models could get, Perez kept asking a smaller, harder question: how does it feel to use one?