Krea 3 Prompting Guide: Get Your Prompts Ready Now

Most people will meet a new image model the same way: open it on launch day, paste in last month's prompts, and decide within ten minutes whether it's any good. That's a bad test. Old prompts carry old habits, and a model built around a new idea gets judged on how well it handles the old one.

Krea 3 is built around a new idea. Krea calls it "our new foundation model, with amazing editing capabilities that build on the aesthetic diversity and personalization we developed for Krea 2." That one sentence is most of what Krea has confirmed so far, and it's enough to start preparing. This Krea 3 prompting guide is about writing prompts, references and edit notes now, on Krea 2, so that launch day is a comparison instead of a scramble.

A word on honesty before we start. Krea hasn't published a model card, a prompt guide or the request format for Krea 3. Everything below that's labeled as confirmed comes from Krea's own words or from the live Krea 2 request on each::labs. Everything else is a working method, and we'll say so.

What We Know About Prompting Krea 3 (and What We Don't)

Three things are confirmed. It's a foundation model from Krea. Editing is its new center. And it builds on the two traits that defined Krea 2: aesthetic diversity (real variety instead of four near-identical outputs) and personalization (steering the look toward yours). We covered the announcement in detail in Krea 3 Is Coming to each::labs.

What isn't public yet: the input fields, how edits are requested, resolution limits, and whether Krea 2's controls carry over unchanged. So the smart move isn't to guess at syntax. It's to get good at the parts Krea says are being carried forward, diversity and personalization, because those are the instincts Krea 3 is meant to keep while you edit.

The good news is that you can practice both today. Krea 2 Large, Krea 2 Medium and Krea 2 Medium Turbo are live on each::labs.

Let Words Carry the Subject and Images Carry the Taste

The single most useful habit with Krea models is a division of labor. Words are good at nouns, actions and arrangement. They're bad at vibe. "Moody", "editorial", "elevated" and "cinematic" have been typed so many times they mean almost nothing to a model, and they mean even less to a model that already has strong taste of its own.

So write the prompt as if you were briefing a photographer over the phone: what's in the frame, where it sits, what light falls on it, what's happening. Then hand the look to references. A good Krea prompt reads almost plain:

A woman in a rust wool coat waits at an empty tram stop at dusk, rain on the glass shelter, a single sodium streetlight behind her, wide shot from across the street.

Notice what's missing: no "masterpiece", no "8k", no adjective pile. The subject is precise, the light has a source, the camera has a position. Everything about grain, palette and mood can come from the images you attach.

Set Creativity on Purpose, Not by Default

Krea 2 on each::labs exposes a creativity setting with four levels, raw, low, medium and high, which controls how loosely the model reads your prompt. Most people leave it on the default and never touch it again. That's leaving the model's signature trait switched to "average".

Treat it as a phase dial. Run high when you're exploring and want the model to wander into directions you wouldn't have drawn. Drop to low once you've picked a direction and need the prompt followed closely. Use raw when you want the least interpretation you can get, which is useful for checking what your words alone produce before references come in.

Why this matters for Krea 3: an editing model has to know when to be faithful and when to be inventive. If you've learned to separate exploring from executing on Krea 2, you'll recognize the same choice in the new model, whatever it ends up being called.

Build a Reference Set That Also Says No

The Krea 2 request takes up to 10 image_style_references, each with a strength from -2 to 2. Positive values pull the output toward a reference. Negative values push it away. That second half is the most underused control in image generation.

Every brand has a look it's trying to avoid: the glossy stock-photo sheen, the over-saturated teal and orange grade, the plastic skin. Describing those in a negative prompt rarely works. Showing them with a negative strength does. A practical reference set looks like this:

{
  "model": "krea-v2-medium-text-to-image",
  "input": {
    "prompt": "A ceramic teapot on a linen tablecloth by an open window, late afternoon light, steam rising",
    "aspect_ratio": "4:5",
    "creativity": "medium",
    "image_style_references": [
      { "image_url": "https://your-cdn.com/house-look-01.jpg", "strength": 1.4 },
      { "image_url": "https://your-cdn.com/house-look-02.jpg", "strength": 1.0 },
      { "image_url": "https://your-cdn.com/too-glossy.jpg", "strength": -1.0 }
    ],
    "seed": 4471
  }
}

Start with two or three positive references that agree with each other and one negative that represents the thing you're tired of seeing. Add more only when a specific quality is missing. A reference set that fights itself gives the model a reconciliation problem, and taste is the first thing it gives up.

Save that set. Krea says the new model builds on Krea 2's personalization, so a curated library of "this" and "not this" images is the most portable thing you can bring to Krea 3.

Write Edit Prompts Like Revision Notes

This is the section where we're reasoning ahead of the specs, so read it as method, not documentation. Krea hasn't said how edits will be requested in Krea 3. What we do know is how edit prompts behave across the editing models already on each::labs, from GPT Image 2.5 to Ideogram 4.5, and the patterns are consistent enough to practice.

Write an edit like a note to a retoucher. One change, stated as an action, plus an explicit list of what must stay:

Change the coat from rust to bottle green. Keep her pose, the shelter, the rain and the streetlight exactly as they are.

Three rules make edit notes work. Use verbs: change, replace, remove, add, relight. Name the protected elements instead of writing "keep everything else", because a concrete list gives the model something to check against. And make one change per pass when precision matters; if three edits land at once and the face drifts, you won't know which one caused it.

Start collecting your real revision notes now. Go through last month's client feedback and write each change as one of these notes. When the new model lands, that list is your test set, and you'll know within an hour whether it saves you work.

Example Prompts to Run on Krea 2 Today

These are written to show off Krea's range and to give you baselines you can rerun later. Run each at high creativity first to see the spread, then at low once you've picked a direction. Save the seeds you like.

An interior with a clear light source:

A narrow bookshop at closing time, one green banker's lamp lit on the counter, stacks of paperbacks leaning against each other, a cat asleep on the till, view from the doorway.

A product shot that leaves room for edits later:

A single amber glass perfume bottle on a slab of pale travertine, hard morning shadow falling to the left, plain warm grey wall, centered, plenty of empty space around the bottle.

A fashion frame with a strong silhouette:

A model in an oversized cobalt raincoat standing on a wet pier, wind pulling the coat sideways, flat overcast light, low angle, gulls in the distance.

A surreal still life, where Krea's range pays off most:

A bowl of lemons where one lemon is a small glowing moon, on a checked tablecloth, deep blue twilight through the window behind.

Then write one edit note for each image (recolor the coat, swap the bottle, move the cat) and keep it with the seed. That's your launch day comparison, ready to go.

Prompting Mistakes That Will Follow You to Krea 3

Stacking style adjectives. If the prompt has more mood words than nouns, the references aren't doing their job. Cut the adjectives and strengthen the reference set.

Testing with one image. Krea's models are built to range. Judging one output is judging a single roll. Generate a spread, then decide.

Using the same creativity for everything. Exploration and final renders are different jobs. If the setting never changes, half your sessions are using the wrong one.

Forgetting the seed. A good image you can't reproduce is a good image you can't edit. Log the seed, the prompt and the reference set together.

Writing edits as wishes. "Make it better" and "more premium" aren't edits. Name the element and the change, or the model will pick one for you.

Wire It Once on each::labs

Whatever the final request for Krea 3 looks like, the each::labs side stays the same: one API, one key, one request shape across the text-to-image catalog. Build your Krea 2 pipeline today and moving to the new model should be a model id and a few field names, not a rebuild. If you chain steps (generate, edit, upscale, animate), sketch the chain now in each::labs flows, so the new model drops into a working pipeline. For a refresher on what makes Krea 2 tick, our Krea 2 review is still the best starting point.

Frequently Asked Questions

Has Krea published official Krea 3 prompting tips?

Not yet. Krea has described Krea 3 as a foundation model with strong editing that builds on Krea 2's aesthetic diversity and personalization, but it hasn't released a prompt guide or request format. We'll update this guide when it does.

Will my Krea 2 prompts work with Krea 3?

Plain, concrete subject prompts and curated reference sets are the most likely to carry over, because those are the parts Krea says the new model builds on. Exact field names may change, so keep prompts, seeds and references logged together and rerun them side by side at launch.

What does the creativity setting do in Krea prompts?

It sets how loosely the model reads your words. On Krea 2 it has four levels: raw, low, medium and high. High explores, low follows closely, and raw shows you what the prompt alone produces.

How many style references can I use with Krea?

Krea 2 on each::labs takes up to 10 style references, each with a strength from -2 to 2. Negative strengths push the image away from a look, which is the fastest way to stop a style you don't want.