Krea (Krea-1)

Krea is the most interesting trajectory story among the image model labs because it started as something that was not a model labModel labA company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google. at all. The company was founded in San Francisco in 2022 by Diego Rodriguez and Victor Perez, two friends who met at the HF0 residency program and shared a frustration with how clunky the existing AI creative tools were. Their original product was a real-time image generation canvas, a browser-based interface where you could type a promptPromptThe text description you provide to a model to specify what you want it to generate. and see images materialize in real time as you typed, rather than submitting a prompt and waiting several seconds for a result. This was built on top of an open-source latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput.-reduction technique called Latent ConsistencyLatent consistencyThe application of consistency-model training in latent space to make diffusion models generate in very few steps. Models, and when it launched in late 2023 it was uniquely magical, nothing else in the field let you iterate on prompts at conversational speed. Real-time generation became Krea's signature feature and the thing that drove early adoption.

Over the next two years, Krea evolved from a single-feature tool into a full aggregatorAggregatorA tool or platform that wraps many image and video models behind one interface, so you switch models without switching apps. platform. They integrated dozens of third-party image and video models, FLUX, Veo, Sora (before its shutdown), Kling, Wan, Hailuo, Runway, Ideogram, Recraft, Nano Banana, and many more, into a unified interface where users could swap between models without switching tools. By 2025 Krea was one of the dominant aggregator platforms for creative AI alongside Flora and OpenArt, with millions of monthly users and a reputation as the place where serious creative professionals built their workflows. The business was working as an aggregator. Then in mid-2025 they did something most aggregators do not: they started training their own models.

The decision to move from aggregation into model training was driven by a specific observation. Krea saw what real users were actually generating, what they refined, what they discarded, what they saved, what they exported, and across millions of users that added up to an unusually clean signal about which outputs humans actually wanted versus which outputs technically scored well on benchmarks. Krea's thesis was that this feedback data could be used to train a model that produced meaningfully better outputs than models trained on scraped web data alone, and that even a small specialist lab could close the quality gap to the frontier labs if they had the right data. The thesis turned out to be correct, and the result is Krea-1.

Krea-1 is not an architecture built from scratch. It is a post-training of Black Forest Labs's FLUX [dev] base model, specifically tuned with a curated dataset and training objective designed to break the model out of what Krea calls the 'AI look.' The AI lookThe AI lookThe polished, oversmooth, conventionally pretty aesthetic that marks an image as machine-made, a result of models regressing toward the average of their training data. is the characteristic signature that makes AI-generated images recognizable at a glance, slightly oversaturated colors, a plasticky quality to skin and surfaces, an aesthetic homogeneity, a tendency to default to certain compositions and lighting setups. Krea observed, correctly, that this look was not inherent to the FLUX architecture but was an artifact of how the base model was trained, the data distribution and loss functions had inadvertently optimized for a particular aesthetic that, once you noticed it, you could not unsee.

The Krea-1 approach was to do extensive post-training on top of FLUX [dev]'s base weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. using a deliberately curated dataset that pushed against the AI look. The training data included reference images drawn from professional photography, editorial magazine work, and selected user outputs from Krea's own platform that human reviewers had rated as visually distinctive. The loss function was adjusted to penalize outputs that hit the signature AI-look patterns. The result is a model that is architecturally compatible with FLUX [dev] (so it works with all existing FLUX LoRAs and tools) but produces noticeably more naturalistic and visually diverse outputs. Black Forest Labs described the model as 'opinionated,' meaning it has a strong aesthetic point of view that can surprise users with how natural the outputs feel.

The open-source version of this work was released as FLUX.1 Krea [dev], which anyone can download from Hugging FaceHugging FaceThe dominant repository for open-source AI models, including all the open-weight image and video models discussed in this document. and run themselves. The closed commercial version is Krea-1, which is the model Krea serves through its own platform and where it continues to evolve through additional post-training cycles. This split between the open research release and the closed commercial evolution is a pattern worth recognizing because it is becoming common across the field as labs figure out how to monetize post-training work that builds on top of open base models.

Krea-1 (mid-2025)

The original closed-model release. Trained as a FLUX [dev] post-training and serving as Krea's house model for users who want the Krea aesthetic. Positioned as an alternative to FLUX [pro] for users whose work benefits from more naturalistic outputs and less of the signature AI look. Accessible only through Krea's platform, no APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. access, no weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. released.

FLUX.1 Krea [dev] (July 2025)

The open-weight companion to Krea-1. Released as a collaboration with Black Forest Labs and available for anyone to download from Hugging FaceHugging FaceThe dominant repository for open-source AI models, including all the open-weight image and video models discussed in this document. under the FLUX [dev] non-commercial license. The model is a drop-in replacement for FLUX [dev] in existing workflows, you can use it with all your existing FLUX LoRAs, in ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org., on fal.ai, anywhere FLUX [dev] works. The aesthetic differences from base FLUX [dev] are subtle enough that casual users may not notice them but trained eyes pick them up immediately.

Krea Realtime (ongoing)

Not a new model but Krea's signature feature, real-time generation where images update as you type or sketch. The current implementation uses a distilled fast variant of the Krea-1 model running at roughly 20 frames per second on Krea's infrastructure. The feature is the primary reason many Krea users choose Krea over competing platforms. As of 2026 it is one of the few production-grade real-time image generation interfaces in the field, though the introduction of distilled models like FLUX.2 [klein] and SANA-Sprint has made the underlying technology more widely available.

Krea Video (2025 onward)

Krea added real-time video generation in 2025 as an extension of the real-time canvas concept. The feature lets users see short video clips update as they iterate on prompts, similar to the real-time image canvas but for the video modality. It is less technically mature than the image side but represents Krea's bet that the real-time interaction pattern that made its image canvas valuable will also be the right user interface for video generation.

Krea's strengths are the aesthetic quality of Krea-1 on naturalistic outputs, the real-time interaction model, the breadth of third-party models accessible through the unified interface, and the platform-plus-model-lab positioning that lets Krea ship creative features that pure model labs cannot match. For professional creatives who want a fluid iterative workflow rather than a submit-and-wait experience, Krea is still the best interface in the field. The ability to swap between thirty or more underlying models without leaving the Krea interface is also genuinely useful for operators who follow the 'route between multiple models' pattern the fal/a16z report describes.

The weaknesses are scale, APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. access, and the question of whether a platform company can sustain frontier-model training investment against the much-larger labs. Krea is significantly smaller than the pure model labs (Black Forest Labs, Midjourney, OpenAI) and significantly smaller than the platform companies (Adobe, Canva, the Chinese internet giants) that could theoretically acquire or out-compete them. Training even a post-training on top of a frontier base model is expensive, and Krea's ability to continue doing this work depends on the aggregatorAggregatorA tool or platform that wraps many image and video models behind one interface, so you switch models without switching apps. platform business continuing to generate enough cash to fund it. API access to Krea-1 is limited compared to the major labs, which constrains Krea's ability to serve enterprise developers and operators building on their model rather than their platform.

Krea is in the most interesting and most precarious strategic position of any lab in this section. They are proof that a platform company can become a model labModel labA company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google. if they have the right data advantage, and the collaboration with Black Forest Labs is a model for how platform-plus-lab partnerships can work without either side getting cannibalized. But Krea is also small enough that the economics of continuing to train frontier-quality models are genuinely challenging, and the question of whether they will continue to invest in Krea-2, Krea-3, and beyond is open. The realistic path forward is probably deeper collaboration with Black Forest Labs and other open-weight labs, where Krea contributes its platform data and post-training expertise and the underlying base models come from larger research operations.

For operators, Krea is best understood as a premium creative tool rather than as a foundational infrastructure decision. If your team works with Krea's interface and Krea-1 fits your aesthetic, the productivity benefits are real and the platform is worth paying for. If you are building a generative imagery product of your own, you should treat Krea as an example of how to build a good creative interface but not as the primary model supplier, because the scale and cost economics favor the larger base-model labs for production infrastructure.

This is a useful pattern to recognize. Increasingly, the work of producing a great image model is being split between the labs that build the base architecture and train the foundation weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. (Black Forest Labs, in this case) and the platforms that do specialized post-training to tune the model for a particular aesthetic or use case (Krea). The open-source release of the post-trained model is a way of contributing back to the ecosystem while also building reputation. Expect to see more of these collaborations.

Krea is best understood as two things at once, and the labels matter. First, it is a real-time canvas and an aggregation layer: one well-designed interface where you paint or type and see the image update almost instantly, and where you can run and chain many other companies' models, FLUX, Wan, Veo, Kling, Hailuo, Ideogram, Seedance, without leaving the workspace. Second, it trains its own models, but the from-scratch story is more recent than the marketing implies. Krea 1, its first self-branded image model in 2025, was tightly coupled to the FLUX ecosystem, and the open FLUX.1 Krea it released with Black Forest Labs was explicitly Black Forest Labs' 12B model post-trained by Krea to match its aesthetic. Krea Realtime 14B, its streaming video model, is a distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.). of Alibaba's Wan 2.1. Only Krea 2, in mid-2026, is credibly a foundation model built from the ground up.

You work in the browser at krea.ai, with a free tier that grants a daily compute-unit allowance and paid tiers reported in 2026 from around nine dollars a month up through a business tier near two hundred, plus an enterprise offering with organizations, zero data retention, and an APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. and MCPMCP (Model Context Protocol)An open standard that lets an AI assistant drive external tools. Applied to ComfyUI, it lets an agent search nodes, assemble a workflow, and run generations on your behalf. integration. Compute-unit and credit pricing can get expensive for heavy video, and lower tiers gate the best video models. The way to use it well is to treat it as the workspace it is: generate and compare across many models side by side, drag in a few images to train a style LoRAStyle LoRAA LoRA trained on images that share an aesthetic but vary in subject, so it can render any prompt in that art style, photographic look, or brand identity. in minutes to lock a consistent look, and move from image to upscale to video to sound in one place, leaning on the real-time canvas for fast iteration.

It is best at real-time interactive image work, aesthetic and style control, quick custom training, and having many models under one opinionated interface. It is weakest as the cheapest option for heavy batch video, and in its structural exposure: because much of the video stack is other labs' models, a change in their access, pricing, or terms flows straight through to Krea, which is the standing risk of any aggregatorAggregatorA tool or platform that wraps many image and video models behind one interface, so you switch models without switching apps., even one that is now shipping its own from-scratch models.