Krea 2: Photoreal Images With Style References

Krea 2 style references let you upload up to 4 images in Krea's app, each with its own strength slider, and Krea 2 makes new scenes in their look: grade, light and composition, not their subjects. Krea's API takes them as image_style_references, with a strength from -2 to 2. On Versely, Krea 2 Medium takes reference images for 3 credits.

The way I test photorealism now is the squint test's opposite: I look closer. Skin at pore scale, the way denim creases at a bent knee, how afternoon light fringes through a glass of iced tea. Most "photoreal" models survive a glance and fail the lean-in — there's a waxiness, a too-even sheen, a suspiciously perfect everything. Krea 2 is the first model in my rotation where the lean-in regularly comes back clean. Its images have the imperfections of photography: sensor-like grain, believable skin, light that behaves like physics instead of rendering.

But raw photorealism is only half of why it's earned a permanent slot in my stack. The other half is style references: feed Krea 2 images that define a look — your past campaign photography, a mood board, a film still — and it generates new content in that visual language. Not similar-ish. In it. For brands with an established photographic identity, that's the difference between "AI images" and "our images, made faster."

Here's how I use Krea 2 in production, where it beats my workhorse models, and where it doesn't. Krea announced it on 12 May 2026 as its "first foundation image model built completely from scratch" (Krea 2), and it ships in three variants: Medium, Large and Medium Turbo (Krea 2 docs).

The photorealism, specifically

What separates Krea 2's realism from the field, from a month of lean-in tests:

• Skin. Texture variation, subsurface warmth, believable specular highlights instead of the airbrushed uniformity that flags AI portraits. People in Krea 2 images look like they have skincare routines, not shaders.
• Optics. Depth of field falls off like real glass; chromatic fringing and vignette appear in trace, plausible amounts. It renders a camera, not just a scene.
• Light behavior. Mixed lighting (window + lamp) resolves into coherent color temperature conflicts, exactly the messy way real interiors photograph.
• Material honesty. Worn edges, dust, fingerprints on steel — the entropy that makes a scene lived-in rather than showroom-dead.

The practical consequence: Krea 2 output survives contexts that scrutinize. Press-adjacent placements, founder-brand lifestyle imagery, anything where "obviously AI" costs credibility. That's a different job from the volume tier — my daily batch work still runs on Flux 1.1 Pro for its consistency — Krea 2 is what I reach for when the image will be looked at rather than scrolled past.

Style references: the brand-lock feature

The workflow that changed my client work: build a style reference set from a brand's existing photography, two to four images that capture its grade, light and composition habits, and generate new scenes conditioned on it.

What transfers, reliably: color grade and palette bias, lighting character (hard/soft, warm/cool, direction habits), contrast curve, compositional density, and that hard-to-name "editorial vs commercial" register. What doesn't transfer: specific subjects, faces, or products — style references govern how it looks, not what's in it. (For locking a recurring subject, that's reference-to-video and character-consistency territory, a different tool for a different problem.)

Rules from doing this across several brand accounts:

1. Curate ruthlessly. Reference images must agree with each other. One inconsistent image (a flash-lit outlier in a natural-light set) muddies the whole conditioning.
2. References set the look; the prompt sets the scene. Don't re-describe the style in words — you'll create conflicts. Prompt subject, action, and framing only, and let the references carry the aesthetic.
3. Version your reference sets. A locked set is a brand asset. Season refreshes get a new versioned set, not casual swaps.
4. Mood boards work as pseudo-brands. For new brands with no photography, build the reference set from curated inspiration first, generate with it, then promote the best outputs into the next reference set. Two cycles in, the brand has a look that never existed before but is entirely consistent.

Style reference settings, from Krea's docs

Krea documents the controls behind the feature, and they explain most of the results above:

Setting	In Krea's app	In Krea's API
Style images	Up to 4	image_style_references
Strength	A slider per image	-2 to 2 per image
Moodboards	Yes, more than 4 images	1 per request, by ID
Output size	Not stated	Only 1K for now

For a single reference, Krea suggests a strength of about 0.6: raise it to make the style dominate, lower it for a lighter touch. Negative values push the output away from a reference's style, and several references blend additively, so Krea suggests starting with strengths that add up to about 1.0 (style transfer docs). Moodboards have to be built in Krea's web app first; the API only points at them.

Krea also says which variant to pick. Medium is the stable, general-purpose model and suits illustration, anime, painting and expressive styles. Large is "more than 2× the size of Medium" and is the one Krea recommends for photorealism, raw textures and the richest output.

Can you use Krea 2 images commercially?

It depends on where you run it.

• In Krea's app: Krea's pricing FAQ says "Basic, Pro, Max, Business, and Enterprise plans include a commercial license that allows you to use all generated content for commercial purposes without restrictions." The Free plan isn't on that list.
• With the open weights: Krea published Krea 2 Raw and Krea 2 Turbo on Hugging Face under the Krea 2 Community License, dated 22 June 2026. Commercial use is allowed only for companies with total annual revenue under $1,000,000. The license also says "You own all Outputs you generate". Bigger companies need a paid license from Krea.
• On Versely: Versely's terms say that, "subject to these Terms and applicable third-party AI provider terms, you may use AI-generated content for personal and commercial purposes."

One common mix-up: FLUX.1 Krea [dev] is a different model, published by Black Forest Labs, and its weights carry the FLUX.1 [dev] Non-Commercial License (model card). That license is not the one Krea 2 uses.

Krea 2 vs the field: routing table

Job	Best pick	Why
Hero lifestyle imagery, press-adjacent	Krea 2	Survives the lean-in
Match an existing photographic identity	Krea 2	Style references
High-volume calendar batches	Krea 3 or Flux 1.1 Pro	Krea 3 is 2 credits and fast
Layout-complete graphics with copy	Seedream 5.0 Pro	Real typography
2K stylistic showpieces	Grok Imagine	Raw resolution mode
Video anchor frames, premium clips	Krea 2 still → I2V	Realism carries into motion

That last row is an underrated pipeline: a Krea 2 frame fed into image-to-video keeps its photographic character through the animation, and premium clips built this way read as filmed rather than generated. It's my default starting point for anything heading into the AI video generator at the commit tier. The full survey of where every image model currently sits is in the latest AI image models roundup and on the live model rankings.

Prompting for the photographic register

Krea 2 responds best to prompts written like a photographer's shot notes, not an artist's vision statement:

• Name the capture context: "shot on a full-frame mirrorless, 50mm f1.8" pulls sharper realism than any quality adjective.
• Embrace imperfection language: "slight motion blur in the background," "uneven window light," "candid mid-gesture" — instructions toward entropy are instructions toward realism.
• Skip the fantasy vocabulary. "Ethereal, stunning, hyper-detailed" push it toward the very renderiness you chose it to avoid.
• Keep one subject clause. Its photorealism concentrates; sprawling multi-subject briefs dilute the magic and raise re-roll rates.

Honest limits

• Resolution. Krea's API renders Krea 2 at 1K only for now, so plan an upscale for print or large hero crops.
• Typography. Middling text rendering — route copy-bearing layouts elsewhere.
• Stylized illustration. It can leave photorealism, but that's buying its weakest skill; dedicated stylized models do it better.
• Style reference drift on distant subjects. Ask a food-photography reference set for a car scene and the transfer thins. Style conditioning works best within shouting distance of the reference domain.

FAQ

What makes Krea 2 more photorealistic than other models?

It renders photography's imperfections — believable skin texture, real lens behavior, mixed-light color conflicts, material wear — rather than an idealized clean scene. Output survives close inspection, which is the actual bar for placements where "obviously AI" costs credibility.

How do Krea 2 style references work?

You supply up to 4 curated images defining a look (in Krea's app), each with a strength slider, and the model generates new scenes in that grade, lighting character and compositional register. References control how images look, prompts control what's in them — don't re-describe the style in words or the two will fight.

Can Krea 2 match my brand's existing photography?

Yes, if your photography is consistent — that's the killer use case. Build a reference set from your best past shoots and new generations inherit the identity. Inconsistent reference sets produce muddy results, so curate hard.

Is Krea 2 or Krea 3 better for style references?

Krea 2. Krea 3, which Krea launched on 15 October 2026, is a fast editing model, and Krea's docs say it "does not yet support the Krea 2 API's style-reference or LoRA controls". Use Krea 3 for quick drafts and edits that keep a product or character consistent; use Krea 2 when a brand look has to carry over. Our Krea 3 launch post covers what it does on Versely.

Can I use Krea 2 images commercially?

On Krea's paid plans, yes: Krea says they include a commercial license. Its free plan isn't listed. The open Krea 2 weights allow commercial use only for companies under $1 million in annual revenue. On Versely, the terms allow commercial use of what you generate, subject to the AI provider's terms.

Should I use Krea 2 or Flux 1.1 Pro?

Both, routed by stakes: Flux 1.1 Pro for volume calendar work where median reliability wins, Krea 2 for hero images, style-matched brand imagery, and stills feeding premium video. Most weeks my split is about 80/20 by count and the reverse by importance.

Krea 2 is live in Versely's text-to-image studio: Krea 2 Medium is 3 credits an image and takes your reference images, and Krea 2 Large is 5 credits. Upload a style set from your best past work and see your brand's look carry over.