Krea released the weights for Krea 2 Turbo yesterday. It is the distilled version of Krea 2, their first in-house image model, and it has been inside the Krea app since May. Now anyone can run it. I ran it through fal.

Every image came back in about two seconds. Not two minutes. Two seconds, round trip, at 1024 pixels. That changes how you use a model. You stop composing prompts and start rolling dice.

Same five prompts I use on everything. Here is what two seconds buys you.

Test 1: Fantasy Castle

Prompt: "Neuschwanstein castle, lightning, pixar style, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight"

Recognizably Neuschwanstein. Sunset on the left, a forked bolt on the right, warm windows. It leans painterly rather than photoreal, which is fair given "pixar style" is in the prompt. The towers are a little soft. For two seconds, I am not complaining.

Test 2: Family Portrait

Prompt: "Family on their laptops while sitting around a Christmas tree with presents underneath and looking worried because they have to finish up work, realistic, 4k"

Four people, four laptops, tree, presents, hands all accounted for. Then it missed the one word that matters. Everybody is smiling. Nobody in this picture is worried about anything. It read the scene and skipped the emotion, which is the oldest failure in this prompt.

Test 3: Evil Mickey Mouse

Prompt: "Evil mickey mouse taking a selfie in Disneyland surrounded by shocked families, realistic, 70s style polariod, 8k"

No refusal. No squirrel. Krea 2 Turbo just drew the mouse. Red jacket, white gloves, Sleeping Beauty Castle in the background, selfie arm out, a crowd of open mouths behind him. It even threw a polka-dot bow on his head, so this is a Mickey and Minnie mashup.

What it did not do is make him evil. He is delighted. The shocked families are reacting to a very happy mascot. Open weights mean no content checker between you and the output, and this is what that looks like. The IP guardrails that every hosted model has built up over the last few years are simply not here.

Test 4: Lionel Messi

Prompt: "Lionel messi wearing his Argentina uniform floating in the air with beams of light behind him posed like Jesus. Sunrise breaking behind him and a soft halo behind his head, in the style of a gothic stained glass window of a church, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight, maximum details"

This is the best image of the set. The likeness is Messi. The sleeve tattoo is there. Number 10 on the shirt, gold boots, sunbeams cutting through a real church interior with stained glass on both walls. The halo is more of a glow than a ring, and the shorts say 19 instead of 10, but the composition is exactly what the prompt asked for.

Test 5: Robot Portrait

Prompt: "rusty robot with bow tie, portrait, 8k, ultra realism, chrome background"

A tin toy. Antenna balls, glowing yellow eyes, a grill of square teeth, a black fabric bow tie on a rusted barrel chest. The chrome drape behind it is convincing. It is charming and it is not "ultra realism." Krea has a house look, and the house look is friendly.

The two-second model

Krea 2 Turbo is not the sharpest model I have tested this year. It missed the emotion in the family prompt and turned the robot into a toy. But it drew a recognizable Messi and a recognizable Mickey without a single refusal, and it did all five in under twelve seconds combined.

Speed is its own kind of quality. When the generation is free and instant, you iterate instead of praying.

The weights are on Hugging Face. The mouse is not going back in the box.