Skip to content
MakerGain
← Back to project

Drag a reference image into the prompt bar: Krea bets twenty million dollars that “taste” can be trained

Edited by MakerGain

Krea turns dragging a reference image into the prompt bar and adjusting a single image’s influence on the generated result with a slider into a concrete product action for style control. Co-founder Victor calls K2 the conservative first self-trained model on the roadmap; it must succeed first, and only then do the next two models have room to take on bigger risks; the team says it spent seven months hand-labeling data and had a twenty-million-dollar GPU cluster run for a year.

For example, you have a reference image in hand: a hand-drawn illustration from an old magazine, with scan grain. You want the model to capture that flavor, not copy the content in the image. What you can normally do is turn the feeling into adjectives—grainy, hand-drawn, vintage print. What you write down and what you have in mind are always a little off.

Krea 2 turns this step into one action: drag the reference image directly into the prompt bar as a style anchor, then slide up and down to adjust how much it affects the generated result. Krea describes this interaction in its own blog On being the underdog[1].

The slider itself is not a big feature, but it puts the hard problem out in the open: to make the model respond to “the style of this image,” adjectives alone are not enough; someone has to mark style in the data. This is also an entry point for understanding why a 37-person company would train its own model.

The trade-off behind a slider

Most frontier labs watch measurable metrics when training models: photorealism, prompt adherence, resolution. Krea says visual style and things like “taste” are mostly formed in the post-training stage, and that is where it concentrated its effort On being the underdog[1]. The same post includes another statement from the company itself: only about two hundred people in the world really know how to do this, and by the post-training stage, it is no longer a model-architecture question but a question of taste.

Krea 2 (officially abbreviated K2) is the company’s first self-trained generative image model, trained for style range and creative variation. The company says it holds up better in directions such as illustration, hand-drawn work, VHS grain, comic animation, and wordmarks, rather than aiming only for photorealism. These are the company’s own claims. In its FAQ, when answering how it differs from Nano Banana and GPT Image, Krea says most large models converge on a single safe style, while K2 is trained for style range—this is self-positioning, not an independent test result, and reading it as a performance win or loss would be a misreading On being the underdog[1].

The cost falls on data preparation: the team spent seven months manually building and labeling its own dataset, and it also wrote the post-training pipeline itself On being the underdog[1]. The blog puts it plainly: it was not that they had to reinvent the wheel, only that the default wheel rolls in a direction they did not want to go. The choice for people building products here is not romantic—if you want the model to recognize “the style of this image,” someone has to label “style” one image at a time.

If a single reference is not enough, the same blog offers another action: first build a mood board and prompt toward the overall aesthetic, rather than stringing together a list of adjectives On being the underdog[1]. The slider adjusts the weight of a single reference image; the mood board governs the overall feel.

“The first model is not allowed to fail”

Co-founder Victor calls K2 the “conservative” first model on the roadmap. The word conservative carries weight: Krea had never trained a model before, and if the first attempt failed, that twenty-million-dollar bet would end there; only if K2 succeeds do the next two models have room to take on more aesthetic and training risk On being the underdog[1].

This is a bet made in sequence, not an all-or-nothing gamble. The same cluster has to run for a year, for twenty million dollars, serving K2 and two other Krea models at the same time On being the underdog[1]. The risk is therefore arranged along a timeline: first prove it can be trained, then talk about whether they dare to take risks. The blog says that once K2 works, the next two models are where they can be “deliberately reckless”: new training norms, new aesthetics, the kind of risk a frontier lab cannot justify because it has to serve a billion users.

The starting point of this path is earlier. The author of this Fast Company report met Victor Perez and Diego Rodriguez in 2023; at the time, the two envisioned not just generative AI media but a set of tools with a craftsman-level sense of control, integrating multiple AI models and blurring the boundaries between image, video, and sound Krea, an Adobe for the AI era[2]. The report sums up Krea’s core promise as “controllability,” and Perez says the original assumption was validated and met with a response in the community. Turning a reference image into a draggable anchor is one concrete shape of that promise.

Image: source Krea’s official redesign post[3], accessed 2026-11-11, original interface preserved.

This official asset comes from the enhancer in the web interface: before-and-after and parameters on the same screen—another form of “controllability” in the enhancer, not K2’s reference-image slider.

Users, funding, and one compute bill

The company says there is a free tier where people can test K2 and other platform features; the “cat riding a bicycle” test is free On being the underdog[1]. Neither the blog nor the Fast Company report gives subscription prices, billing periods, or credit rollover, so there is no way to judge how many people actually pay.

The scale of money and people is known: Krea’s official blog says the company has raised a cumulative $83 million and has 37 people On being the underdog[1]; the Fast Company report says the $47 million B round was led by Bain Capital Ventures, after an earlier A round led by Andreessen Horowitz, for a cumulative $83 million and a $500 million valuation Krea, an Adobe for the AI era[2]. The same report describes the scale when the B round closed as “two years later, 20 million users,” but the body gives no user definition, precise counting date, or channel breakdown—the number can show how visible it has become, not how much revenue it makes.

On the cost side, only one figure is known: that twenty-million-dollar GPU cluster that has to run for a year. It corresponds to compute investment, not a monthly inference bill; public information does not break out model procurement, inference, storage, or personnel, and the profit structure remains unclear.

Whose style is it

One detail does not read like marketing language: Krea says content users upload or generate still belongs to the users, and users can choose not to let their work be used to train Krea models On being the underdog[1]. In the same blog post, the company says it is experimenting with using AI itself to identify creators who have a “measurable style influence on a given generation,” hoping to explore a sustainable layer of royalties before the industry settles On being the underdog[1].

There is no shipped version of this, and the wording is “experimenting” and “hoping”; whether it can be calculated clearly, and who would be paid once it is calculated, both remain open. But it and the slider are two ends of the same question: if style can be pointed out and adjusted by sliding, then “whose style is this” will sooner or later need a ledger. Krea put the royalties experiment in the same piece of text as the release of its self-trained model, which at least shows it knows that bill is coming.

What Krea is betting on is that taste can be trained, productized, and perhaps credited; and the starting point of that chain is to first make style into an object that can be dragged, slid, and marked with weight. Whether a style looks right is ultimately verified on the user’s own image, not in any blog post.

Continue reading

AI video
Putting pictures on the beat: neural frames turns synchronization into a business

neural frames handles the synchronization involved in setting music to pictures: it separates elements such as drums and vocals so generated visuals land on the beat. Founder Nicolai Klemke began with general text-to-video, then followed advice to narrow the product to music videos because the product lacked a well-defined use case. The small team buys models and APIs externally, concentrating its work on separating songs and passing parameters to the models.

AI images
A hundred thousand signups at zero marketing cost: Supermeme finds the template first

Supermeme reverses meme-making: users describe a situation first, then the system finds templates by meaning and drafts editable captions. It takes over template browsing and the first caption; people still judge tone and potential offense. In that interview, Sanjeev said the three founders had never met offline, collaborating on WhatsApp and twice-monthly Zoom calls, with largely reactive product discovery and instinct-led building decisions. Strictly speaking, they had yet to go all in and each had other projects. Sanjeev remained a product manager until November 2022, then explored Supermeme full-time while consulting to pay the bills.

AI companionsAI imagesAI audio
One reply counts as one, a photo as eight: Embraces.AI breaks companionship into messages and sells it

Embraces.AI prices companionship by message: one reply counts as 1 message, and one Picture this image counts as 8. Companions have cross-session memory, send multiple bubbles in a row, and initiate contact; users can edit memories and export chats. Before signing up, there are 40 free messages; the three paid tiers start at $12, and $5 tops up 250 messages. TrustMRR recorded on 2026-11-14 all-time revenue of $48,193 and 426 active subscriptions.

AI imagesAI videoAI audio
59 Million Views but Almost No Revenue: After Publishing the Entire Project, Overseas Signups Increased to 6.5x on the Day

Zohar Vanunu used his self-built Kolbo.AI to make a “biblical phone imagery” AI series, which drew 59 million views and more than 200,000 followers in 30 days, but by his own account it brought in almost no revenue; after publishing the complete production tutorial, overseas signups increased to 6.5x on the day, and in three weeks 3,285 people accounted for 78% of new signups. The product puts more than 100 models, along with chat, image, video, music, and voice, into the same project and charges by credits.