Krea 3 vs Krea 2: Two Models, One Logo, No Overlap

The fastest way to understand the relationship between Krea 2 and Krea 3 is to try to write a single prompt that both can serve. You cannot, and not because of quality. Krea 2 is a text-to-image foundation model — you give it words, it gives you an image, and it was published on 12 May 2026 with its weights later released openly as Krea-2-Raw and Krea-2-Turbo. Krea 3, released on 15 October 2026, cannot do that at all. It is an editor: it takes an image or a set of references and changes what you tell it to change. Five months separate them and there is no line in the API where one substitutes for the other. automatic failover when a provider degrades a routing DSL that lets you express

One makes images, the other changes them

The distinction is easier to trust when it comes from the artefacts rather than the marketing.

• Krea 2's inference script takes a prompt, steps, guidance, resolution, seed and image count. Those are generation parameters. There is no flag that accepts an input image, and Krea's public repository holds no editing checkpoint. The platform around it sells editing, but it does that by calling other labs' editing models, not Krea 2.

• Krea 3's launch page is a catalogue of edits. Swap a sofa, keep the rug and the reading woman. Change a cup to cherry red. Re-skin a billboard for a different brand. Take a line drawing and a flat six-colour palette and materialise it as one architectural photograph. It renders natively at 2K with 4K for print, and every example on the page starts from something that already exists.

• The reference counts point in opposite directions. Krea 2's style control is about feeding the model a look — style references and moodboards, where a curated collection of images steers aesthetics. Krea 3's references are operational: up to six per edit, each one a person, product or style that must survive into the output unchanged.

• Speed is measured differently because the work is different. Krea 2's page claims outputs in 15 seconds or less. Krea 3 claims under eight seconds end to end for a 2K edit. The Krea 3 figure is a round trip through an edit; the Krea 2 figure is a generation from nothing.

If you have a blank canvas, Krea 2 is the model. If you have an approved asset that a client has already signed off and you need one thing about it changed without touching the rest, Krea 2 cannot help you and Krea 3 exists precisely for that.

Why a Krea 2 user should care about Krea 3 at all

Because the two compose, and the composition is the interesting part.

Krea's own launch material for Krea 3 says the model "has Krea 2's eye for style, so it works on paintings, vector art and ink drawings and keeps them in their own medium." That is a claim about inherited aesthetic judgement, and it is the kind of claim a vendor makes when it wants continuity between generations. Take it as a description of intent rather than a measurable property.

What is measurable is the pairing. Krea 2 generates a hero image and its style variants; Krea 3 then adapts that hero across formats, languages and colourways without regenerating the parts that were approved. The two-step pipeline is the one Krea itself demonstrates — a packshot becomes a lifestyle scene, an ad extends into a series, a poster localises into another language — and none of those steps is a generation problem.

• Generate with Krea 2, adapt with Krea 3. The expensive creative decisions happen once, in the generation step, and the cheap repetition happens in the edit step.

• Reference images carry across. Krea 3 takes up to six references per edit; a character or product established for Krea 2's style work can be handed to Krea 3 as a consistency reference instead of re-described in every prompt.

• Speed is what makes the pairing practical. A sub-eight-second edit is short enough that iterating on a variant costs less attention than writing the prompt for it.

The price is the same number and the same shape

Both are billed per image, and the Krea 3 rate happens to land inside Krea 2's published range — which makes the hosted-versus-download question worth spelling out.

• Krea 3 — $0.03 per image at 1K, $0.06 at 2K. Krea's FAQ states this is list price, not a promotional rate, and says the company expects it to fall.

• Krea 2, downloaded — no per-image price at all. The RAW and Turbo checkpoints are on the Hub under the Krea 2 Community Licence. You pay in hardware and in the licence's revenue condition, not per call.

• Krea 2, hosted — Artificial Analysis lists three tiers at $15, $30 and $60 per 1,000 images for Medium Turbo, Medium and Large respectively. Those are served endpoints, not the checkpoints.

• Comparing a downloaded checkpoint to a hosted service is comparing different things. $60 per 1,000 hosted images against $60 per 1,000 Krea 3 edits — the numbers rhyme, but one buys a generation and the other buys an edit.

That last bullet is the practical reason the two models do not cannibalise each other commercially. They sit in the same price band because per-image pricing across the image industry has settled into that band, not because they are competing for the same call.

The open-weights line does not extend to Krea 3

Krea 2's open release is the most consequential difference between the two, and it is worth being precise about what "open" bought you.

• Two checkpoints, two jobs. RAW is the undistilled base, meant for fine-tuning, post-training and LoRA training. Turbo is an eight-step distilled checkpoint for fast inference. The documented workflow is to train on RAW and run on Turbo, which is why the community's LoRA ecosystem is built that way.

• The licence is not permissive. The Krea 2 Community Licence permits commercial use while you and your affiliates stay under one million US dollars in trailing-twelve-month revenue across all sources. Above that, you need an enterprise licence from Krea before further commercial use. Redistribution carries the agreement, a "Krea 2 [Your Model Name]" prefix and the attribution notice. Delaware law governs.

• Krea 3 is closed. There are no checkpoints for it. The launch FAQ answers the self-hosting question — "Yes, on enterprise plans. Krea 3 runs on your own GPUs in your cloud" — which is a licensing conversation, not a download.

• The API reference registers Krea 3 as a beta endpoint. The model is served at api.krea.ai/generate/image/krea/krea-3/ and labelled "Krea 3 Beta" in Krea's own documentation, next to three Krea 2 endpoints that carry no beta label. Krea's launch post opens by saying it is opening the beta today.

So the two models sit on opposite sides of the openness line as well as opposite sides of the generation/editing line, and anyone standardising on Krea should know which side of each line their workload lives on.

Where each one stands on the scoreboards

This is the section where the two models are least comparable and where a vendor's silence is the loudest signal.

Krea 2 has an independent record, and it is a large one. On Artificial Analysis's text-to-image leaderboard, read 16 October 2026, Krea 2 Large sits at 1,034 Elo from 14,090 samples, at $60 per 1,000 images; Krea 2 Medium Turbo at 1,018 Elo from 5,672 samples; Krea 2 Medium at 1,013 Elo from 14,045 samples. That board carries 102 models and puts the leader at 1,217, so Krea 2 Large is a middle-of-the-table model with a fourteen-thousand-vote sample behind it — a real measurement, not a flattering one. There is no Krea 2 entry on the image-editing board at all, which is the same fact as "Krea 2 does not edit," expressed as a leaderboard.

Krea 3 has nothing. No Artificial Analysis entry, no arena result, no neutral run of any kind — it launched on 15 October 2026 and this research was collected the following day. The numbers attached to it so far are Krea's own: under eight seconds, $0.06 at 2K, six references, trains on a product or a person. Krea's speed demo publishes its own methodology — median request-to-result time of a twelve-image batch including queue time, measured on Krea's platform from 24 September to 1 October 2026 — and notes that one comparison price on the same page is a draft estimate pending validation. That is unusually candid for a launch page, and it is still a vendor measuring itself.

The consequence for a buyer is straightforward. Krea 2 is a model you can place on a leaderboard and in a procurement document today. Krea 3 is a model you can place in a pilot.

Running both without running two vendors

The awkward part of a two-model pipeline is not the models; it is the plumbing. Two API surfaces, two authentication schemes, two sets of rate limits, two things that can be down at the same time in different ways.

OrcaRouter carries neither Krea model today, so nothing here changes if you are already calling Krea directly. What we do offer is the layer that makes a multi-model pipeline a configuration rather than an integration project: one endpoint across 200-plus models, automatic failover when a provider degrades, a routing DSL that lets you express "generate here, then edit there" as one call, and 0% markup with provider list price passed straight through — so if Krea follows through on the price reduction its FAQ anticipates, that reduction is live on our side the same day the vendor posts it.

For a pipeline that mixes a vendor's generation model with the same vendor's editing model, the value is not in savings on any single call. It is in the fact that a new checkpoint or a new price does not require a code change to adopt.

What to actually do

Use Krea 2 if the input is a prompt. Use Krea 3 if the input is an asset. The overlap people keep looking for is a naming artefact — Krea numbered a new product in a sequence, which reads like a successor and behaves like a sibling.

Three things worth watching, in the order they are likely to be answered. Whether Krea 3 gets an independent score, since it currently has a price and a latency claim but no third-party measurement. Whether Krea 2's open weights get a generation-three counterpart, because Krea's roadmap language so far points at resolution and compute efficiency rather than at open releases. And whether the "$0.06 is list price" line holds, because that single sentence is what makes the volume arithmetic in Krea's own launch material work.