Vibeleaderboard
Krea

ai lab

Krea

Krea matters because it treats latency, continuous visual feedback, and creative control as model requirements rather than interface polish added after training. Its move from a multi-model product into an original image-model program and adapted real-time video research creates a useful test of whether a specialist company can turn a creator feedback loop into differentiated model development without matching the compute scale of the largest labs.2,7,3,8

Official site ↗Models in the Atlas →

Profile

Overview

A creative interface company

Krea is a San Francisco creative AI company founded in 2022 by Victor Perez and Diego Rodriguez Prado, who met as students in Barcelona. The company began with an interaction thesis rather than a proprietary foundation model: creators needed a visual workspace that could hide model fragmentation and respond to sketches, prompts, images, and live inputs. Its early real-time canvas became the basis for a broader product spanning image generation, video, editing, enhancement, and training.1,2,5

Real-time generation becomes a workflow

The company grew through a workflow layer that orchestrated both outside and internally adapted models. Krea's real-time interface continuously regenerated an image as a user changed a prompt, moved shapes, painted, or streamed a screen or camera. By April 2025, Krea said it had reached more than 20 million users and increased revenue twentyfold in fourteen months with a team of eight. Krea announced $83 million in aggregate funding across a $3 million pre-seed and seed round, a $33 million Series A, and a $47 million Series B; reporting placed its post-money valuation at about $500 million.2,5,1

From model orchestration to model development

Krea then moved toward greater vertical integration. It released a Krea-tuned FLUX checkpoint, Krea Realtime 14B, which it distilled from Wan 2.1 14B for real-time autoregressive video generation, and Realtime Edit for direct painting and live visual transformation. Krea 2 extended that strategy into an original image foundation-model family, with Raw and Turbo open-weight variants under a custom community license and hosted Medium and Large product variants.6,7,8,9

Krea 2 tests vertical integration

The Krea 2 technical report describes data curation, a transformer architecture, multi-stage training, distributed infrastructure, and distillation for faster generation. Independent reporting frames the effort as a roughly $20 million bet by a smaller company to compete through different aesthetic choices rather than maximum scale. Krea therefore now has two connected advantages to test: an original model program and a creator interface that can generate direct product feedback. Its newest benchmark and adoption claims still need more independent reproduction before they carry the same evidentiary weight as the product's established reach.8,3,4

Notable contributions

  1. 01A continuously updating generative canvasKrea made sketches, prompts, shapes, webcam input, and screen input part of a continuously updating visual canvas, shifting generation from repeated submissions toward direct manipulation.2,7
  2. 02Real-time video distillation at 14B scaleKrea Realtime 14B distilled Wan 2.1 14B into an autoregressive video model using Self-Forcing, adding memory and inference techniques for streamed generation and mid-generation prompt changes.6
  3. 03Krea 2 image foundation modelsKrea 2 established an original image-model stack with documented data, architecture, training, and distillation choices, plus Raw and Turbo weights under a custom community license.8,9,4
  4. 04A workflow-to-model feedback loopKrea connected its proprietary model development to a creative product used across image, video, enhancement, and editing workflows rather than releasing the model as an isolated endpoint.1,3,8