Krea Release Notes

Introduction

In this technical report we introduce Krea 2: a series of foundation models designed for both wide aesthetic diversity and user creative control. We present: (1) our principles for data curation; (2) the model architectures; (3) our multi-stage training pipeline; (4) our distributed infrastructure; and (5) future work.

Sangwu Lee, Erwann Millon, Le Zhuo, Matthew Newton, Andrei Filatov, Naga Sai Abhinay Devarinti, Dazhi Zhong, Avram Djordjevic, Gabriel Menezes, Will Beddow, Titus Ebbecke, Mihai Petrescu, Owen Fahey, Gian Saß, Felix Gil, Albert Salgueda, Victor Perez

The model weights and inference are released under a permissive license.

Introduction

Over the past few years, image generation has seen remarkable progress. Diffusion and flow-matching models can generate high-resolution images, produce sharp photorealism and stable structure, render dense text, encode broad world knowledge, and follow user prompts in precise detail. These improvements have been driven by several interacting factors including scalable transformers architectures, improved captioning and text encoders, better latent representations, and pipelined post-training techniques. Yet as the field has optimized for reliability on these capabilities, many systems have converged toward a narrow set of default aesthetics. While effective production tools, this makes them less effective as engines for creative exploration, where users often need to search across styles, moods, compositions and visual directions rather than receive a single polished default.

To address these limitations, we present Krea 2, a series of foundation models focused on creative exploration. Krea 2’s models are built on the belief that image generation should be an exploratory medium: expressive enough to span many aesthetics, and controllable enough for creators to navigate them.

We built a large-scale data infrastructure and distributed training framework from scratch to curate a comprehensive pretraining dataset with broad world knowledge and style coverage.

Using this infrastructure, we train expressive models through a multi-stage pipeline spanning pretraining, midtraining, supervised finetuning (SFT), preference optimization, and reinforcement learning (RL), with each stage designed to progressively refine the model’s output distribution. We develop a simple yet performant diffusion transformer (DiT) architecture through thorough ablations. Our model incorporates several components that accelerate convergence, including iREPA, improved VAEs, and Qwen3-VL. We also integrate several architectural improvements, including grouped-query attention (GQA), sigmoid-gated attention, lightweight timestep modulation, and multilayer feature aggregation for text-encoder features, which together improve training stability and efficiency.

A strong base model is only useful if users can reliably reach the parts of its distribution they care about. In training, the model learns from rich, carefully constructed captions that describe images with dense visual detail. In practice, user inputs are often shorter, more ambiguous, and shaped by many different habits of expression. Some users describe a scene in natural language; others gesture toward a mood, a style, or a reference image. This creates a gap between the model’s learned conditioning space and the way creative intent is expressed at inference time.

To reduce this gap, we build two systems that make Krea 2 more exploratory and steerable from both text and image inputs: a prompt expander and a style-reference system. The prompt expander maps simple or underspecified user prompts into richer visual directions without overwriting the user’s intent. It is trained through a two-stage SFT and RL pipeline on top of open-source LLMs, where the objective is not only to improve image quality, but also to encourage creative variation and controllable exploration. Complementing this textual interface, the style-reference system lets users express visual intent through images when words are insufficient. It allows users to inject the style or mood of one or more reference images with minimal content leakage, while providing fine-grained control over style strength and weighted style mixing.

Together, these components define Krea 2 as a foundation model for exploratory generation. Instead of optimizing only for a single polished default, Krea 2 is designed to expose a broad visual space and give users practical ways to move through it, using both text and image-based control. Krea 2 is among the top 10 models on the the Artificial Analysis leaderboard for text-to-image, and scores 2nd place among models from independent labs. Krea 2 serves as a comprehensive baseline and enables a creative generative experience while maintaining competitive performance.

[The report continues with detailed sections on Data Curation Principles, Captioning, Architecture, Optimization, Training Pipeline, Prompt Expansion, Style Reference System, Training Infrastructure, Systems Infrastructure, Observability, Lessons Learned, Filesystem and Checkpointing, Data Infrastructure, Discussion and Future Work, and Conclusion.]

Conclusion

We presented Krea 2, our first foundation-model series built for creative exploration. In this technical report, we shared the infrastructure, data systems, training pipeline, and research choices behind the model family. We believe foundation image models are still in their early stages, and we look forward to sharing more of our research.

Join us at Krea

We're building the future of creative AI. Come work with us.

By the Krea Team.

Krea 2 now gives you more control before the image is generated. Generative Sliders let you shape the model’s creative direction while keeping your original prompt intact, so you can explore different outcomes without rewriting the same idea over and over.

Open the controls panel in the Krea 2 image tool to adjust Intensity, Complexity, and Movement, choose a Creativity mode, randomize settings for exploration, or reset everything back to neutral.

Move a single setting and watch Krea 2 shift the result while the prompt stays steady.

What Generative Sliders Control

The Generative Sliders panel gives you direct control over four parts of the generation process:

Control | What it changes
Intensity | Moves from quieter, more muted images to more strongly stylized images.
Complexity | Moves from simpler, more minimal compositions to denser, more visually complex images.
Movement | Moves from more static images to more dynamic compositions with stronger camera angle, pose, and motion energy.
Creativity | Controls how much Krea expands or interprets your prompt before generation.

Intensity, Complexity, and Movement are centered Generative Sliders. A value of 0 is neutral. Moving left reduces that quality; moving right increases it.

See Generative Sliders In Action

Keep the prompt steady and move one slider at a time. These examples show how the Movement and Complexity sliders can change the generated image while preserving the core idea.

“cute poodle, city street, colorful illustration, playful composition”
Movement slider · -100, -50, 0, 70, 100

“illustrated cafe table with a croissant, coffee cup, and oversized vase of wild flowers”
Complexity slider · -100, -50, 0, 70, 100

Creativity Modes

Use Raw when your prompt is already precise. Use Low or Medium when you want Krea 2 to add useful visual detail without drifting too far. Use High when you want the model to explore the idea more aggressively.

Try Random Mode

Random mode changes the slider settings each time you submit a generation.
It is useful when you want to explore the shape of an idea quickly. Turn it on, generate a few results, and look for the direction that feels closest. Once you find something interesting, turn Random mode off and fine-tune the individual sliders from there.

A Few Starting Points

For clean design, icons, and editorial illustration, try lowering Complexity and keeping Movement near neutral.
For cinematic images, fashion, worlds, or expressive character work, try raising Intensity and adding a small amount of Movement.
For precise prompts, set Creativity to Raw or Low. For loose prompts, set Creativity to Medium or High and let Krea 2 build more of the visual language around your idea.

Try Generative Sliders

Open Krea 2 and use Generative Sliders to shape your next image generation.

By the Krea Team. Last updated June 2026.

Introducing Krea 2 Turbo.

Generate high-quality images in just 2 seconds with a faster Krea 2 experience, compatible with style references, moodboards, and LoRAs. You can try it for free in Krea.

What is Krea 2 Turbo?

Krea 2 Turbo is a faster version of Krea 2. It is designed to keep as many of the original model’s capabilities as possible while running multiple times faster.

That means you can stay inside the Krea 2 workflow, use the same creative controls, and get visual feedback much sooner. Turbo is especially useful when you want to explore an idea quickly before spending more time refining the final direction.

Built for faster creative loops

Image generation is not only about the final output. A lot of the creative work happens while you are testing prompts, changing references, comparing compositions, and deciding which direction is worth pursuing.

Krea 2 Turbo makes that loop faster. It is a good fit for rapid ideation, early concept work, moodboard exploration, product mood passes, art direction, and prompt experimentation.

Compatible with Krea workflows

Krea 2 Turbo works with style references, moodboards, and LoRAs, so you can use it in the same creative workflows you already use with Krea 2.

Because Turbo is optimized for speed, Krea 2 Medium and Krea 2 Large are still the stronger choices when exact style adherence or final-image polish matters most. Turbo is for the moments when you want to move quickly, see more options, and keep momentum.

Try Krea 2 Turbo

Generate high-quality images in just 2 seconds with Krea 2 Turbo.

By the Krea Team.

Last updated May 2026.

Krea Train lets creators teach a model a repeatable visual language: a brand style, a character, a product, a material study, or a personal subject. We are bringing that workflow closer to Krea 2.

Krea 2 LoRA training is starting as a closed beta inside Train. During the beta, selected creators and teams can train LoRAs for Krea 2 Medium and Krea 2 Large, use them in Image, and choose whether a trained style stays private or becomes a public LoRA other people can discover and reuse.

Open Train

Build a custom style, character, or object model from your own dataset.

What is a public LoRA?

A LoRA is a lightweight custom model trained on a focused dataset. Instead of rewriting a long prompt every time you want a specific look, object, or character, you train that concept once and reuse it as a style layer in generation.

A public LoRA is the same idea, made shareable. You can keep trained LoRAs private for personal or workspace use, or make selected LoRAs public so collaborators, communities, and future Krea workflows can build from the same visual identity.

Why Krea 2 LoRAs matter

Krea 2 is built for stronger taste control, sharper image quality, and a more flexible relationship between prompts and visual references. LoRA training adds another layer: persistent memory for the things you care about most.

For a designer, that might be a product line or campaign style. For a filmmaker, it might be a character or world treatment. For an architecture or interior team, it might be a material palette, rendering language, or office-specific presentation style. The point is not just to generate one good image, but to make a visual system reusable.

What you can train

Krea Train supports several common training intents:

• Styles: teach Krea a visual treatment, illustration language, material quality, lighting approach, or brand look.
• Characters: keep a person or fictional character more consistent across poses, scenes, and prompts.
• Objects: preserve a product, prop, furniture piece, or design object across new contexts.

Krea 2 LoRA training supports Style, Character, and Object modes. Krea 2 Medium and Krea 2 Large are wired as image training models, with model-specific training settings handled by the Train workflow.

Existing Train examples for styles, objects, and characters

How the beta works

The flow is intentionally simple: upload a focused set of images, choose what you are training, review or add captions when needed, and start the training job. When the LoRA is ready, it appears in Train with your other styles so you can use it in generation, reuse the dataset, update the cover, or manage its visibility.

The minimum dataset is small, but quality matters more than quantity. A cleaner set of images that repeats the same subject, style, or object usually teaches the model more clearly than a large, mixed dataset. Image limits can vary by plan, and Train will show the active limit inside the upload flow.

Private by default, public when you choose

Training is most useful when it respects the way creative teams actually work. Some LoRAs should stay private: client concepts, internal brand systems, unreleased products, or personal subjects. Others are better when shared: a public art style, a reusable community preset, or a model that helps a team create from the same visual vocabulary.

That is why Train keeps visibility as an explicit choice. A LoRA can stay private, or it can become public when you decide it is ready.

Access and availability

Train is currently available to paid Krea users with access to model training. Krea 2 LoRA training is a narrower beta: eligible users and workspaces will see Krea 2 Medium and Krea 2 Large in the Train model picker. We are using the beta to test quality, captions, training defaults, safety checks, queue behavior, and how public LoRAs should feel once more creators can use them.

As access expands, the goal is straightforward: make custom Krea 2 models easier to train, easier to reuse, and easier to share without turning the workflow into a technical chore.

Train your own visual model

Upload a focused dataset, train a style, character, or object, and use it across Krea generation tools.

Open Train

By the Krea Team. Last updated May 2026.

Moodboard Gallery just got a major expansion

We are rolling out a larger Moodboard Gallery for Krea 2 with thousands of new moodboards to explore and generate from. The goal is simple: faster style discovery and less setup before you start creating.

What is new

• Thousands of new moodboards are now available in the gallery.
• You can browse references and generate with Krea 2 from those moodboards.
• We are previewing two new selection modes: Random and Auto.

Random and Auto (preview)

These two new modes are designed to speed up early exploration:

• Random picks a moodboard for you when you want a quick unexpected direction.
• Auto selects moodboards automatically based on your generation context.

Both modes are in preview and may continue to evolve as we collect feedback.

See the update

A quick look at the expanded gallery and the new Random/Auto preview modes.

Why this matters

Moodboards are one of the fastest ways to shape visual direction in Krea 2. Expanding the gallery and automating moodboard selection helps you get to a strong starting point faster, especially during ideation.

Try Moodboard Gallery

Explore the expanded gallery and test Random and Auto in Krea 2.

Source update: @krea_ai on X.

Krea 2 is our first foundation image model

Krea 2 is our first foundation image model, trained completely from scratch to give users full control over the look, feel, and creative direction of every image they create.

Starting today, you can access Krea 2 through our API partners, such as Fal, Comfy, or Runware. We’re also launching with Nous Research, offering access to Hermes since day 0, and you can find support for Claude, OpenClaw, and Codex in our docs.

Here’s what it brings to the table:

• Aesthetic diversity – Krea 2 can generate an incredibly wide range of styles, with particular strength in artistic and expressive directions that other models tend to flatten out.
• Style transfer – Krea 2 ships with the most advanced style transfer system available. It can extract the style of one or many reference images and apply it to your outputs with precision.
• Creativity – this is an actual parameter you can tune with Krea 2. At high values, it will make the model to add aesthetic depth, interpretation, and visual richness to whatever you give it.
• Moodboards – you can pass in an entire moodboard of dozens of images and Krea 2 will understand the overall creative direction, producing outputs that align with it.

Find hundreds of examples with their respective prompts clicking the link below:

Browse the Krea 2 gallery

Hundreds of examples with the prompts that produced them.

Open the gallery

We built Krea 2 for creatives who want real control over what AI produces. We work closely with teams across creative studios, marketing and advertising agencies, fashion, gaming, and architecture — continually taking their feedback back to our research lab to push the model further for every use case. It’s a model that’s actively evolving with the people who use it.

It’s also built for exploration. You don’t always come in with a fully formed idea, and Krea 2 is designed to be a creative collaborator in those moments — something that can take a loose direction and give you back visuals that spark the next idea, rather than forcing you to specify every detail upfront.

Model variants

The Krea 2 API offers two model variants:

Krea 2 Medium and Krea 2 Large. Both support the same parameters and feature set — the differences come down to model size, training stage, and the character of the outputs they produce.

If you’re not sure which to start with, try Krea 2 Medium — it’s faster, cheaper, and handles the broadest range of use cases reliably.

Krea 2 Medium

A smaller, faster, and more cost-efficient model with extensive post-training applied. This makes its outputs especially stable and consistent across generations. Medium is particularly strong on illustration, anime, painting, and other expressive or artistic styles.

Krea 2 Large

More than twice the size of Medium, with a softer post-training applied – which gives its outputs a rawer, more textured, and flexible character. Large is the more powerful model overall – at its best, it produces results Medium can’t match. It handles expressive and artistic styles well, and shines on photorealism and “raw” looks like motion blurs, grains, or low dynamic ranges.

Features & Examples

We compiled this list of examples that you can use to test the capabilities of the model.

Aesthetic diversity

Most image models handle complex prompts well, but they tend to fall short on style — defaulting to something sharp, polished, and “safe”. That narrows the creative range you can actually get out of them.

With Krea 2, we put as much effort into how the model understands style as we did into how it understands prompts. The result is a model that can render a wide range of aesthetics – expressive, raw, niche, experimental – without falling into AI looks.

The following are a set of examples created with Krea 2 showcasing a variety of aesthetics:

“illustration of a cat riding a bicycle”

Model: Krea 2 Large

“heavily blurred dark silhouette of a head and shoulders, glowing reddish-orange center on the obscured face, thin neon lime green fire-shaped sparkles above, cool blue-grey background, heavy vintage film grain, lo-fi grunge aesthetic, dusty scratched texture”

Model: Krea 2 Large

“Close-up portrait of a woman, retro 90s anime style, large vivid blue eyes, flushed cheeks, glossy red lips, silver glitter around her eyes, single brown curl on the nose, wearing a glossy black metallic helmet with rivets, bold blue reflections, dark background, heavy film grain, saturated colors, sharp specular highlights”

Model: Krea 2 Large

Style Transfer

Krea 2 ships with the most powerful style transfer system on the market. Pass in a single reference image or combine several, and Krea 2 will extract the style and apply it to your output — letting you decide how strongly each reference shapes the final image.

The following were generated using style references:

Creativity

Creativity is an actual parameter that you can tune for Krea 2. When a prompt is short or vague, the model can fill in the missing pieces — style, composition, camera angle, color palette — adding visual depth where you haven’t specified it. The creativity parameter controls how far that expansion goes.

Set it to high and the expansion gets stronger — the model takes more creative liberty with style, mood, and aesthetics. Set it to low and the model sticks more closely to what’s in your prompt. Set it to raw and the model performs no expansion at all, rendering only what you’ve explicitly described.

Moodboards

A moodboard is the most precise way to set a visual direction with Krea 2. Moodboards are built from a group of images that share an overall creative direction — palette, texture, style, mood, or composition. You can create one in Krea and reference it when you run inference through this API.

The following examples show how activating a moodboard can transform the output of even very simple prompts.

API Setup

Create an API token from your Krea account settings. Krea 2 has two versions of the model available in the API:

Endpoints

Krea 2 Medium

https://api.krea.ai/generate/image/krea/krea-2/medium

Krea 2 Large

https://api.krea.ai/generate/image/krea/krea-2/large

Parameters

Both versions of Krea 2 support the following parameters.

• prompt (string): Required text prompt
• aspect_ratio (string): 1:1, 4:3, 3:2, 16:9, 2.35:1, 4:5, 2:3, 9:16
• resolution (string): Currently just 1K
• seed (number): Optional seed
• creativity (enum): raw, low, medium, high
• image_style_references (array): Optional style references
• moodboards (array): Optional moodboards, currently max 1
• styles (coming soon) (array): Optional trained styles (LoRAs)

Fetch completed jobs

Completed jobs can be fetched from:

https://api.krea.ai/jobs/{job_id}. You can also pass an X-Webhook-URL header to receive a POST when the job completes.

More detailed API docs are available at docs.krea.ai/developers/introduction. It covers things like the job lifecycle, webhook integrations, generating api tokens, etc.

API Examples

Below you can find a set of short guides that will show you how to use all the different features that come with Krea 2:

• Generate an image
• Use style transfer
• Adjust creative thinking
• Use moodboards

Pricing

Model type | Text-to-image | Style references | Moodboards
Medium | $0.030 | $0.035 | $0.040
Large | $0.060 | $0.065 | $0.070

Note: combining moodboards with style references will not increase the price per generation, i.e., it will be $0.04 for Medium and $0.07 for Large.

Our launch partners include Runware, Fal, Nous Research, and ComfyUI.

By the Krea Team. May 2026.

Krea 2 is our first foundation model, built completely from scratch and focused on aesthetics and creative control.

When you make an image with AI there are really two questions:
what you want in the image, and
how you want it to look.
Most models are great at the first. They handle complex prompts without breaking a sweat.
But when it comes to style they default to something polished, safe, and a little generic — the AI look.

We’ve talked before about that AI look and the barriers it creates when you’re trying to get something expressive or artistic out of a model. For Krea 2 we wanted to take it head-on.
We put as much work into how an image looks as into the model itself.
The goal was something capable of rendering almost any style — from the grainiest film photography to the cleanest studio shot, cinematic stills, illustration, digital painting, and weirder experimental territory.
And just as importantly, something you can steer, not just prompt at.

The reason matters.
The kinds of products and features we’ve been able to build at Krea have always been limited by the models we had access to.
Starting our own research lab means we can finally make the technology that lets us build creative tools that treat AI as an actual creative medium — raw, flexible, unopinionated, uncut.
Something you can break if you want to, that doesn’t quietly box you into the same handful of looks.

This post is a walkthrough of the three things that make that possible in Krea 2: open-ended exploration in the image tool, style references, and moodboards.

Watch the full deep dive

14 minutes of Krea 2 in action.

Start vague. Let the model explore.

Krea 2 doesn’t expect a fully-formed idea.
You can come into the image tool with something as bare as a cat riding a bicycle and just hit generate two or three times.
What comes back isn’t four variations on a single safe interpretation — it’s the model thinking through completely different ways that concept could be rendered.
Painting. Old VHS photography. Minimal illustration. Abstract shapes with their own internal logic.
The first round of generations becomes a kind of mood-sweep — a way to see what kinds of images even live inside this prompt, before you commit to any one of them.

This is a different posture than what most image models train you into.
You don’t have to come in with a detailed brief.
You can use the first three generations as research, find the direction you like, and then narrow from there.
It’s closer to the way an art director works — start broad, then home in — than the way most AI tools force you to write.

Tighten the prompt just a little — say, a cat riding a bicycle, retro cartoon illustration — and you still get diversity, but the cluster narrows.
All of them feel like retro cartoons; the variety lives inside that style.
Some are more complex, some more minimal, but they’re all unmistakably retro cartoon.
The model isn’t picking a single canonical interpretation of “retro cartoon” and printing it four times; it’s exploring inside that style space.

The same thing works across radically different modes.
Switch the prompt to a cat riding a bicycle, dreamy cinematic scene at 16:9 and you get a whole spectrum of takes on that vibe — different lighting, different camera languages, different moods, all sitting comfortably under the same “dreamy cinematic” umbrella.

And if you want to push the model toward something less polished, you can.
Extremely grainy lo-fi VHS still gives you exactly that — low resolution, chromatic aberration, the kind of texture most models actively try to clean away.
This is where you can feel the unopinionated part of Krea 2 most clearly.
The model isn’t holding back, isn’t quietly steering you toward “tasteful,” isn’t softening the edges of the look you asked for.

Style references: dial the look up and down

Once you’ve found a look you like — whether it came out of one of these exploratory rounds or you brought it in from somewhere else — style references let you carry it forward into completely different prompts.
We spent as much time on the style transfer system as we did on the foundation model itself.

The way it works: you drag any image into the prompt box, into the area marked add as style transfer, and then you write whatever prompt you want.
Under the hood, the system extracts the stylistic components of that image — color, texture, composition cues, the painted-or-photographic feel of it — and applies them to your new generation.
So if you liked the painting style of one of those cat-on-bicycle outputs and want to see what it does with a completely different subject, you can.

The most important knob here is strength.
Strength controls how hard the style pushes against the base model.
At 50% you get a balanced blend — recognizably stylized, but the prompt is still firmly in charge.
The interesting part is what happens at the extremes.

Drop the slider to 20% and the same prompt with the same reference produces something much more realistic.
You can still see hints of the reference: a touch of the color palette, a faint painted quality.
But the base model has the upper hand and it pulls the image toward photography.
This is useful when you want a style to inform a generation rather than dominate it — when you want a whiff of the reference instead of the whole thing.

Push it the other way to 80% and the relationship inverts.
Now the reference is in charge.
You get a full painting in the same palette, the same brush strokes, the same compositional language as the reference image.
Push it even further and the model starts hunting for places to put the reference’s color — sometimes overriding the subject itself to make the style fit.

You can see this in the outputs where the model tries so hard to transfer the bike’s color into the horse that it starts painting parts of the horse itself in that color.
That kind of “breaking” is informative; it tells you exactly where the style transfer is trying to land.

The other thing you can do — and this is where style references really start to feel like a creative instrument — is combine them.
You can stack up to four references at once, each with its own strength slider.
Add a darker knight reference at 70% on top of an illustration at 50% and you start seeing the interesting part: stylistic components from both images coexisting in the same output.
The chrome and the darkness from the knight, the painted aberrations and the looser hand from the illustration.
The horse comes out partially chrome, partially painted, with backgrounds and lighting borrowed from both worlds.

Re-balance the same two — illustration to 75%, knight down to 57% — and the result tilts toward the illustration: more of the painted artifacts, only certain parts of the horse rendered in chrome instead of the whole thing.
You can literally feel each slider pulling on a different aspect of the output.

Add a third reference into the mix — a line illustration at 80% on top of the painting at 60% and the knight at 44% — and the whole process starts feeling less like prompting and more like turning physical knobs on a creative tool.
You don’t have to write any new text.
You can move sliders, swap references, raise one and drop another, and watch the output rearrange itself in response.

For us this is the most honest answer to the question “how do you give an AI tool real creative control”: you give the user variables that map to visual qualities and let them play.

Moodboards: a richer kind of reference

Style references are precise.
You hand the model a specific image, it extracts the style, you control how hard that style pushes.
Moodboards are the other new thing in Krea 2 and they work on a different axis. Two important differences:

• No four-image cap. You can throw as many images at a moodboard as you want. The whole point is that a moodboard is a set, not a single reference, and the system reasons over the set.
• The algorithm is doing more than style transfer. Moodboards use style transfer under the hood, yes, but on top of that they run custom LLMs and clustering across your images. They pick up on concepts, recurring characters, expressions, compositions, atmosphere — the actual mood of the set, not just its visual style.

The workflow is simple.
Open the sidebar in the image tool, click moodboards, create a new board, drop in a batch of images, and click analyze.
After a moment you get three columns back:

• Taste profile. A high-level description of the kinds of things the algorithms found in your moodboard. Useful both as a sanity check and as a kind of mirror — sometimes you see your own taste described back at you in a way you wouldn’t have written yourself.
• Keywords. The set of style tags the system will use under the hood every time you generate with this board.
• Avoids. Things the system will actively steer away from when you generate with this board. This is one of the more underrated bits — being able to encode what you don’t want is often more useful than being able to encode what you do.

Then you generate as usual.
Prompt a frog against a board of colorful illustration and you get frogs that pull the palette and the line work — and sometimes go further, picking up incidental details like little stars from the references and weaving them into the scene.

The output isn’t a “frog in the moodboard’s style” so much as “what would happen if a frog existed inside this universe.”
Sometimes that turns into the frog quietly watching its own bubbles like a confused tourist. That’s the point.

We’ve also shipped a set of presets so you can play with the system before you build your own.
Each preset is a hand-curated moodboard with its own taste profile, keywords, and avoids — you can swap them in and out and instantly see the same prompt rendered through completely different aesthetic universes.
Retro web is one of our favorites.

Apply it to the same frog prompt and you get pixelated, slightly chaotic 3D-collage territory — bits of late-90s web aesthetic colliding with rendered objects and stickers.
Sometimes the outputs lean realistic-with-a-vibe, like a collage that almost makes sense; sometimes they go fully pipe-and-3D-blob. Both are correct.

Expressive marker swings the same prompt into character territory.
The reference set is a bunch of expressive, slightly goofy illustrated characters, and the system keeps that mood consistent across whatever you generate with it.
The frog comes out looking a little funny, a little stylized, very obviously a character and not just an object in a style.

This is the part of moodboards that style references can’t quite do on their own: transferring not just look but attitude.

The mental model that’s worked best for us: where style references are about transferring a look with precision, moodboards are about handing the model a vibe and seeing what it does with it.

Try it

Krea 2 gives you a lot of room to explore and a lot of aesthetic control.
Exploration mode for figuring out what you want.
Style references for carrying a specific look across prompts and dialing it in with strength.
Moodboards for handing the model a whole creative universe and letting it generate inside that universe.

They’re designed to be used together — start with exploration, refine with style references, lock in an aesthetic universe with a moodboard — but each one stands on its own too.

The best way to feel the difference is to open the image tool and start pushing it.

Open the image tool
Try Krea 2, style references, and moodboards.

How it works

Krea 2 is built around style control: not just what should appear in an image, but how it should feel. Style references and moodboards already let you guide Krea 2 with visual examples. LoRA training takes that one step further.

Starting today, we’re rolling out Krea 2 LoRA training in beta — our most powerful fine-tuning system to date. Train Krea 2 on your own specific style, object, or character with precision, then use it directly in the Image tool.

Krea 2 LoRAs are available for Max and Business subscribers while we test in beta.

LoRAs aren’t new to Krea — we’ve supported them on Flux, Qwen, and Wan for a while. What’s new is the quality. Krea 2 LoRAs hold style, character, and material identity at a level we haven’t seen from any of the earlier backbones.

Three steps from a folder of images to a reusable Krea 2 LoRA.

1. Open the trainer

Head to /train, click Train new LoRA, and hit Next step.

2. Add images, pick the model, set steps

Drop in a set of images (at least three) that share the style, character, or object you want to capture.

On the right sidebar, choose Krea 2 Medium or Krea 2 Large — Large trains stronger, Medium trains faster. Adjust Steps if you want the LoRA to learn more aggressively; the default is tuned to work for most datasets.

Krea auto-captions your dataset before training. Captions tell the trainer what to learn and what to ignore, so a one-off background or prop doesn’t accidentally become part of the LoRA. You can edit captions before kicking off the run.

3. Use it in the Image tool

When training finishes, your LoRA appears in your Train library with preview images. Open it from there, or hit the LoRA button while generating in the Image tool with Krea 2 selected.

You can dial strength up or down, and combine multiple LoRAs — stack a character with a style, or layer two visual languages on top of each other.

When to reach for a LoRA

Style references are great when you want to guide a single generation. A LoRA is better when you want a visual direction to become reusable — a painterly look applied across subjects, a character placed into new scenes, a product or brand language you can keep generating from without rebuilding the direction every time.

Krea 2 LoRAs are especially useful when the thing you care about isn’t easily described in a prompt: a texture, a recurring face, a brand object, a sketch style, a lighting habit, a strange material.

Made with a Krea 2 LoRA

beccu.studio
@beccu.studio
“A custom creature LoRA trained on Krea 2 — same identity across wildly different scenes and moods.”

A few tips

Use a focused dataset.
For a style, keep images visually consistent but varied enough that the model learns the style instead of memorizing one composition. For a character, include different angles, expressions, lighting, and backgrounds.

Captions matter.
If an image has a detail you don’t want the LoRA to learn, call it out in the caption. Auto-captions are a good start; a quick manual pass usually makes the LoRA noticeably better.

Start with the defaults.
Krea shows the estimated compute before you submit, and the default Steps count works for most datasets. Bump it up only when you want the LoRA to learn more strongly.

Keep your first LoRA narrow.
One clean character, one style, or one product identity beats trying to teach everything at once.

Private by default

Trained LoRAs are private when created. Keep them to yourself, share with collaborators, or make them discoverable later. If you make a style public, other Krea users can generate with it — but your original training images stay private.

Train a Krea 2 LoRA

Upload a set of images and turn a style or character into a reusable Krea 2 model.

Start training

Frequently asked questions

By the Krea Team. Last updated May 2026.

Mood boards vs. style references

Mood boards are one of the new tools that ship with Krea 2, and honestly — they’re our favorite.

A quick refresher on style references: you pick up to four images, and the model transfers their style onto whatever you prompt. It’s focused, surgical, and gives you a lot of control over how that transfer happens.

Mood boards are a different animal.

Two big differences:

• More than four images. You can pile in as many as you want.
• A more complex system underneath. Mood boards sometimes use style references behind the scenes, but they also run custom LLMs and clustering methods to read everything about the images you upload — not just style. Concepts, expressions, characters, the overall vibe — it all gets factored in.

The result: a tool that goes further than style transfer. It pulls the whole world of your reference set into every generation.

How to use mood boards

Open the sidebar in Krea — Mood Boards is the second item. Click in, hit create new mood board, drag and drop your images.

Before you can generate with it, you need to analyze the board. That’s the step where all those algorithms run and Krea builds a profile of your references.

Once analysis is done, you get three columns:

• Taste profile — a high-level description of what the system found in your board.
• Keywords — style tags that get applied under the hood every time you generate with this board.
• Avoids — things the system will actively steer away from.

Hit generate mood board to drop back into the image tool with the board selected. Now anything you prompt will be filtered through that whole aesthetic.

What it actually does

Prompt something simple — say, “a frog” — and you’ll see the mood board doing more than copying colors. If your references have stars floating around, your frog might end up in a starry scene. If the expressions in your references are goofy, your frog comes out looking a little dummy too.

That’s the whole idea. Style references transfer style. Mood boards transfer everything — concepts, ideas, palette, mood — and generate with all of it at once.

Sometimes it locks onto one dominant style. Sometimes it blends. And sometimes it pushes into something a little crazier that still belongs to the world of your references. That last part is what makes it so good for exploration.

One prompt, three mood boards

Same prompt — “a frog” — generated through three different boards. Same subject, completely different worlds.

Presets

Krea ships with a bunch of curated mood board presets. A few to try:

• Retro web — pixelated UI components, low-res 3D vibes, collage-y aesthetics.
• Expressive marker — hand-drawn characters with funny, loose energy.
• Film noir — photographic, cinematic, a little grainy.

Same prompt, different preset, completely different output. It’s the fastest way to feel how much a mood board changes a generation.

Why we love it

Mood boards give you aesthetic control without locking you into a single look. They leave room for the model to surprise you — within the world you defined. For anyone who likes to explore visually, it’s easily one of the most powerful tools in Krea 2.

Try mood boards in Krea 2

Build your own board, analyze it, and generate.

By the Krea Team. Published May 12, 2026.

This is Krea 2, our first foundation model built completely from scratch and focused on aesthetics and creative control. When creating an image with AI, there are two main pieces that you need to define: what do you want in that image, and how do you want that image to look?

Most image models are great at the first part. They can handle really complex prompts with no problem. But when it comes to style, they often want to give you something that looks good. Something polished, safe, and nice. Nothing too crazy. We have talked about the AI look in the past, and about the barriers it creates when you try to get expressive or artistic styles out of these models.

So for Krea 2, our focus has been on how you want the image to look. We wanted something capable of rendering pretty much any style: from the grainiest film photography to the cleanest studio shot, cinematic stills, different kinds of illustrations, complex digital paintings, and any kind of experimental direction.

To get there, we put as much effort into building a style transfer system as we put into creating the foundation model itself. You can pass an image into this system, and Krea 2 will understand its style components and transfer them into the output. It also lets you control how much the style should influence the final image, and it lets you combine multiple styles together.

This is one of the main ideas behind Krea 2: style should not be a vague prompt word. It should be something you can guide, mix, strengthen, reduce, and push. In the image tool, you can use style references to shape the output directly. You can add multiple references, combine them, and adjust how strongly they influence the image.

If you want the model to stay closer to a reference, you can do that. If you want the reference to be a lighter direction rather than the whole image, you can do that too. The model also includes controls for how much variation you want across a batch. You can keep a set of images more cohesive, or push it toward a wider spread of palettes, details, and style choices.

The goal is not just to generate a good image. The goal is to give you room to explore a visual direction. If there is one thing we want to do at Krea, it is to build creative tools. The kinds of products and features that we have been able to build up until now have always been limited by the kinds of models we had access to.

With the start of this research lab, we expect to create technology that allows us to build creative tools that treat AI as an actual creative medium. Something that feels raw, flexible, unopinionated, and unconstrained. Something that you can break if you want to. Something that does not limit the types of visuals you can create with it.

We cannot wait to see what you do with Krea 2, and we are getting ready to ship many new updates and features over the following weeks.

Try Krea 2

Generate images with our most expressive model yet.

By the Krea Team. Published April 2, 2026.

Seedance 2.0 is available on Krea. It’s the strongest AI video model we’ve tested for motion coherence, camera control, and clip-level stability.

Seedance 2.0 is now available on all paid plans. Try it in the video tool.

What it does

Full-clip coherence. Lighting, textures, and spatial relationships hold from first frame to last. No mid-shot drift, no warping.

Camera pans keep direction. Tracking shots maintain depth. A slow reveal looks intentional, not approximated.

Characters stay recognizable throughout the clip. Faces, clothing, proportions persist across time instead of getting rebuilt frame to frame.

The output is stable enough to edit, composite, and drop into production pipelines. Close-ups and macro shots hold up under scrutiny.

How to try it

Open the video tool, pick Seedance 2.0 from the model selector, write your prompt, and generate. Same workflow you already know.

Availability

Seedance 2.0 is available on all paid plans. Pick any paid tier and start generating. Check out our plans.

Try Seedance 2.0

Best-in-class motion coherence and camera control.

By the Krea Team. Published March 26, 2026.

Krea Edit now supports annotations

Mark up multiple regions of an image, write a separate prompt for each one, and generate all the changes in a single pass.

Editing an image region by region used to mean working through them sequentially. Select an area, describe the change, generate, wait for the result, then move to the next one. Five things to fix meant five separate rounds.

With annotations, you mark up everything first, then generate once. Draw rectangles around every area that needs work, write what you want for each one, and the model processes all regions together.

How it works

Open an image in Krea Edit and start drawing rectangles on the canvas. Each region gets a numbered badge in its own color so you can tell them apart. A card appears anchored below each rectangle where you type your prompt.

You can describe something completely different for each region. Swap a background in one, change an outfit in another, add an object somewhere else. Each prompt is independent.

Annotations can also carry reference images. If you want a region to match a particular look, drop the reference right into that region’s card. Upload from your files or pick from your existing assets. The model uses it as a visual guide for that area only.

The sidebar lists every annotation with its color badge and prompt text. Click any entry to jump to that region on the canvas.

One pass

When you generate, the model handles each prompt on its own while keeping untouched areas intact. One generation, all your edits.

For single changes, the brush and selection tools work the same as before. Annotations are for when you look at an image and see three or four things to fix at once.

Try Annotations

Edit multiple regions in a single generation.

By the Krea Team. Published March 18, 2026.

Starting a workflow

Krea’s Node Agent turns a sentence into a runnable creative workflow. You type what you want to make. It reads your canvas, plans the pipeline, wires the nodes, and runs the job.

Agentic AI has mostly lived in developer tools until now. Here’s what it looks like when it’s built for creative work instead.

You write something like “cinematic video of a futuristic city at sunset” and the agent starts by reading your canvas. Whatever’s already there, it factors in: existing nodes, connections, outputs from earlier runs. If you have a style node from a previous session, it reuses that instead of creating a new one.

Most AI tools treat every request as a blank slate. The Node Agent doesn’t. It sees your full canvas state, not just the words you typed.

Plan, then build

The agent shows you a plan before it touches anything. For that futuristic city, the plan might be four stages: prompt refinement, image generation on K1 or Flux, enhancement, then video. You can swap models or drop stages before it starts building.

Once you approve, nodes appear on your canvas one layer at a time. Each one wires into the last as it’s placed.

Validation and cost

Before the job runs, the agent walks the full graph. If a parameter is missing, it fills it in. If an output is feeding into an incompatible input, it either fixes the connection or adds a conversion node between them.

Then it shows you compute cost, broken down per node. Nothing runs until you say go.

Iterate, don’t restart

Results show up right on your canvas.

When you change something, the agent only reruns what’s downstream of the edit. Swap a prompt at the top of a ten-node chain and the first nine cached results stay put. Only the affected nodes reprocess. You’re iterating on a piece of the pipeline, not rebuilding the whole thing every time.

Under the hood

Planning and reasoning come from frontier LLMs, including Anthropic’s models. Generation runs on Krea’s own models and whatever else is available on the platform: K1, Flux, Imagen, video, enhancement. The agent picks based on what you asked for and what’s already on the canvas.

Most agentic AI tools target developers and data teams. This one is for people who make things for a living. A ten-node creative pipeline shouldn’t require you to learn node graphs. You describe what you want and adjust what comes back.

Try the Node Agent

Describe what you want and watch it build.

By the Krea Team. Published March 9, 2026.

Krea Edit is a rebuilt editing tool that gives you direct control over specific parts of an image. Select a region, describe what you want, and the AI changes just that area while leaving everything else untouched.

Editing AI images used to mean regenerating the whole thing and hoping the part you liked stayed the same. With Krea Edit, you point at the thing you want to change and describe the change. The rest of the image holds.

Region editing

Draw a rectangle, paint with the brush, or click an object to auto-select it. Once you have a selection, type what you want that area to become. The model fills it in, matching the lighting and style of the surrounding image.

You can also move objects around. Grab something in the scene, drag it to a new position, and the tool inpaints the gap it left behind. Expand the borders of an image by dragging outward and describing what should fill the new space.

Creative controls

Beyond region edits, Krea Edit includes a set of tools that change the character of an entire image. Relight a scene by placing up to four virtual lights, each with its own color, brightness, and position. Adjust camera angles to render new perspectives of the same subject. Apply a color palette to shift the mood of a shot.

There’s also a camera lens simulator. Pick a camera body, a lens type, a focal length, and an aperture. The model re-renders the image as if it were shot with that setup. A portrait on a Petzval lens at f/1.4 looks different from the same portrait on a tilt-shift at f/8.

Colorize turns grayscale images into color. You can drop color hints onto specific areas to guide it, or let the model decide on its own.

How to open it

You can open Krea Edit from any image you’ve generated. Look for the edit button in the lower-right corner of the screen. Click it and the editor opens as an overlay with the image already loaded, ready to work on. You can also go directly to krea.ai/edit and drop in any image.

Every edit stacks. You can make a regional change, relight the result, shift the palette, and keep going. The generation history strip at the bottom tracks each step so you can go back to any earlier version.

The idea

Generating an image is one step. Getting it to look exactly the way you want is the rest. Krea Edit is built around that second part: the fine-grained, iterative work that turns a good generation into a finished piece. The most capable tools should be the simplest to use, and that applies to editing just as much as it does to generating.

Try Krea Edit

Fine-grained AI editing for every image.

By the Krea Team. Published March 5, 2026.

Drop any image into Krea and it writes the prompt for you.

A detailed description appears in the prompt box, ready to generate with.

Finding a reference image is the easy part. Turning it into the right words is where things slow down, especially the lighting, the style, the details that make it work. Instead of guessing, drop the image in and let Krea describe it for you. The prompt appears ready to use, and you keep moving.

See it in action

How it works

Drag any image onto the image tool and a “Create Prompt” zone appears. Drop it there. A description starts streaming into the prompt box word by word, like watching someone type.

What comes out is a dense prompt, 30 to 100 words, tuned for models like Flux and K1. Use it directly, edit it, or add it to a prompt you’ve been working on. Change your mind halfway and cancel. The text stops where it is.

Works with images dragged from the web too. No need to save the file first.

What it sees

The tool reads an image the way a good prompt engineer would. It looks at medium and style first, since those have the biggest effect on what the model generates. Then subject, lighting, composition, color, environment, and any text visible in the image.

A generic captioner would tell you “a photo of a person standing outside.” This tool gives you “editorial portrait, soft directional light, shallow depth of field, muted earth tones, outdoor cafe, overcast sky.” Try generating from both and you’ll see why that matters.

Streaming

The prompt streams as it’s generated, five words at a time. You can read it while it’s still being written. Sometimes the first few words are enough to know you want to cancel and try a different reference.

We went back and forth on the streaming speed. 60ms between each batch of words is where we landed. Faster than that and the text just appears all at once, which defeats the point. We wanted you to be able to read along and decide whether to keep going.

Prompt engineering is a real skill, and people have gotten good at it. But a tool that does it for you in seconds, from any reference, without learning the vocabulary first, should just be a drag and drop. We’ve always believed the most capable tools should be the simplest to use, and this is another step in that direction.

Try It Now

Drop any image and get a generation-ready prompt.