Krea on X

Approval is flat and personal rather than comparative: people post pictures, short films and 3D references they made and call the quality "insane", say it "really works", or describe using it all the time in a workflow. Krea 3 is the version they name — 2k generations, edits in 8 seconds, a clean interface, one render to get the shot right.

7 people name precise editing as what won them: changing one element without repainting the rest, bounded regions, hex colour control, and a one-prompt removal of a bystander from a photo. 6 praise reaching Krea through other assistants, including an MCP inside Claude and ChatGPT and a Grok bot that renders images in chat, one of them a five-minute film written, shot and scored that way. 6 praise Krea 3 for e-commerce: product angles, consistency across a catalog, and the speed of it as a budget saver — level with the assistant group.

5 are interested but not converted, saying they will try it or wondering how it holds up on everyday prompts. 4 doubt the praise: that Elo scores measure preference rather than accuracy, that results fall short of the trailers, or that the open-source community was dropped.

The complaints are few and unrelated, and each is raised by about the same number of people.

Two people describe cancelling as deliberately hard: one counts eight clicks to cancel a subscription and calls it "an adobe level shit", naming a growth service that uses dark patterns; the other files it as a dark pattern too.

Two worry that the details that make an image usable in a store are the ones that drift. A warped logo or a slightly wrong colour makes the photo untrustworthy, and in e-commerce one of them says that is "the whole game" — the image becomes unusable rather than merely imperfect.

Two raise the cost of steady use rather than of a single render: a usage limit in the way of regular work, and a hope that API latency and pricing stay reasonable once people wire it into production. One of them puts it plainly — the models are good, "except for the usage limit".

Nano Banana is weighed on edits and on scores. One user says nano banana makes them start over when a minor detail is wrong, while Krea 3 handles perspective changes and repeated edits without spoiling the image; another rates Krea 3 Ultra over nano banana pro for edits; a third notes Krea's Elo score landing one point behind nano banana 2.1, "remarkably close for an independent lab".

Qwen comes up as the model people are leaving or measuring against. One used Qwen Image 2.1 for university work and was disappointed at how it handled complex structures, and wants a better open model so the work stays reproducible; another asks whether Krea can edit at all and wants to see how it fares against Qwen.

With GPT Image, the comparison is uneasy: one person asks how Krea compares with it alongside nano banana and Grok, and another says that unless Krea is less censored than Image 2.5, it is hard to bother.

Midjourney gets one mention: first place, with Krea added to the list for what it does above the image model.

No question is asked by more people than another; each of the four comes from about the same number.

The practical questions come first. People ask which image models are inside Krea, or which one another user runs there, and whether Krea is a paid subscription. When it runs from someone's own account inside another app, one user finds it unclear where the units are drawn from: "where the units come out of they still don't say".

Others test the limit of the output. One asks how it does with complex structures — towers, skyscrapers, complex bridges — and another whether it can generate above 2k.

A last pair want the method, not the verdict: the "secret sauce" to a logo-to-brand-identity technique, asked by a designer building portfolio work, and how much a noise overlay changed a result.

The open-weights ask is the only one with a group behind it. People want the weights released so they can run the model themselves, one naming a 16GB VRAM GPU as the target, another waiting for Ultra on Hugging Face. The reasons given are practical: a university researcher wants work that stays reproducible and a model that copes with complex structures; others simply ask for open source, or to be woken when the weights drop.

Two want finer control over small fixes. One works on character art and says fixing small details still takes forever, calling editing "the biggest bottleneck in my workflow"; the other asks to change only a character's motion while the face, clothes and scene stay put, rather than generating many attempts.

The rest are single requests: don't use bounding boxes ("PLEASE, DO NOT USE BBOXES"), don't forget the Turbo version, support JSON as an option because it is powerful to work with, and a plain request for early access.