Krea 2 prompting guide, tested

Write Krea 2 prompts as dense, plain-English description: who or what, doing what, wearing what, where, in what light, and shot how. Krea’s technical report says it plainly: “Dense prompts reliably produce better image-generation results.” We tested prompt length, format, adjectives, negation and text rendering on Krea 2 Turbo, three seeds each, and the results are below.

Every render here was made on SirenRender’s free Quick Generate tile: Krea 2 Turbo at 8 steps and CFG 1, 960×1536, with no prompt rewriting in between.

1. Length: more words, more control

On Krea’s own site a separate prompt-expansion model rewrites short prompts into rich captions before the image model sees them. The report describes that expander as its own component, and the open weights take your prompt as written, which is how SirenRender and ComfyUI run them. So the detail has to come from you.

The six-word prompt gave a believable photo, but the model picked everything: who she is, what she wears, how she stands. The one-sentence prompt followed every detail and came out looking like a stock photo. The 115-word prompt, which also names the light, the lens and the film, gave the most finished-looking pictures, and nearly the same frame on all three seeds. That’s the trade: every detail you write is one less thing the seed gets to decide.

Krea’s own model card uses a fox walking in the snow as its example. On Krea 2 Turbo it gave nearly the same picture on all four seeds: a fox walking straight at the camera, filling the frame. Short prompts don’t mean variety; they mean the model’s default. One sentence that adds a birch forest, overcast light and a telephoto lens pulled the camera back into a wildlife shot, with the fox small among the trees. The full paragraph brought it close again and fixed the exact moment: one paw lifted, snow on its coat, framed by birch trunks. One detail in it never rendered: the fox’s breath in the cold air.

2. Prose, tags or JSON

Krea’s report says its users prompt in “natural language, tags, detailed JSON, bounding boxes, instructions, visual guidelines, and Markdown”, and that training captions were reformatted “into a variety of lengths and formats”. We wrote the same details three ways.

All three followed every detail: the copper hair, the cable-knit sweater, the tulips, the striped awnings. Prose and JSON both gave the candid moment the prompt describes, her lifting the tulips mid-laugh. The comma-tag version turned it into a posed, front-facing portrait. If you think in fields, JSON works fine on Krea 2. If you want a moment rather than a pose, write sentences.

3. Camera words beat beauty words

“Stunning, flawless, gorgeous supermodel with a perfect symmetrical face” plus “ultra realistic, 8k, masterpiece, best quality” gave a centered, unsmiling headshot on every seed. It looks expensive and generic. Describing the person and what she’s doing, then naming a lens and a film stock, gave a picture with something happening in it. The quality words added nothing the camera words didn’t do better.

4. Say what’s there, not what isn’t

A short negation worked here. “No people, no cars” came back as an empty street on all three seeds, the same as the version that calls the street empty and deserted. Krea 2 reads prompts through a Qwen3-VL language model, and it handled a two-item “no” without trouble.

Long lists are a different story. In an earlier batch of 40 Krea 2 Turbo renders made for another project, a closing line of “no pumpkin, no jack-o’-lantern, no ghost, no bat…” put those exact things into 12 of the 40 pictures. Naming an object puts it in play. Keep a negation to one or two ordinary things, and for anything specific, describe what is there instead.

5. Text in the image: put it in quotes

Krea 2 spelled “FRESH TULIPS” correctly on all three seeds, in the cream letters on a dark green board the prompt asked for. Put the exact words in double quotes, then say where they go and what they’re painted or printed on.

A template to start from

See the Krea 2 camera prompts page for how angles, lenses, film stocks and light change the same scene, and the Krea 2 prompt library for finished prompts with all four seeds shown.