← All organizations

AI creative tools and image foundation models

Krea

Krea builds AI tools for creating and editing images, video and 3D assets. Its workspace brings together more than 50 models, custom model training and enhancement tools for creative teams in advertising, architecture, fashion and film. Krea Nodes lets users connect generation, enhancement and style transfer into reusable workflows, while enterprise controls manage model access and spending. Krea also acquired Wand and released Krea iPad, combining native brushes with real-time AI.

Founded in 2022 by CEO Victor Perez and CTO Diego Rodriguez, Krea grew from their work on interfaces for artistic interaction with AI. Its own Krea 2 foundation-model family focuses on creative exploration. A prompt expander turns sparse descriptions into richer visual directions; a style-reference system lets users control and mix reference-image aesthetics while limiting the transfer of their subject matter. Krea released open weights for its Raw and Turbo variants in 2026.

Krea reported over 20 million users in April 2025. In April 2025, the company reported $83 million in cumulative funding. That financing included a $47 million Series B led by Bain Capital Ventures, at an approximately $500 million reported post-money valuation.

www.krea.ai

Start here

  1. Perceptual Evaluations: Evals for Aesthetics — Diego Rodriguez, Krea.ai

    Start here for an explanation of how human sensory limits shape perceptual compression, providing context for judging generative-media evaluations.

    Diego RodriguezAI Engineer World's Fair 2025

  2. Training Krea 2 - What matters in generative model training.

    Read for concrete dataset methods: combining OCR with vision-language-model captions and using Wikipedia concept rankings to improve coverage.

    Sangwu LeeAI Engineer World's Fair 2026

  3. Infra behind Krea 2 - How to train and serve at scale

    Read for how language-model research ideas were adapted to diffusion transformers trained across InfiniBand-connected GPU clusters.

    Gabriel Jorge MenezesAI Engineer World's Fair 2026

Messages from the stage

When benchmarks miss visual quality

Diego Rodriguez critiques FID and CLIP-style proxies, showing how visually subtle compression artifacts can distort scores and arguing for evaluations that learn subjective human preferences.

Training data for aesthetic diversity

Sangwu Lee describes preserving unconventional styles while avoiding synthetic training images. His account connects captioning, deduplication, filtering, and concept coverage to building Krea 2's visual knowledge.

Making GPU capacity usable

Gabriel Jorge Menezes discusses thermal throttling and misleading utilization figures alongside tensor-core and networking metrics. Kueue gang scheduling and Kubernetes workload priorities complete the operational picture.

Affiliations reflect each recorded session, not necessarily current employment. The shared HF0 showcase's source summary flags discrepancies in presenter listings and session matching.

Company sources · checked 2026-08-27