Krea 2 prompt enhancer, tested
Krea publishes the instructions it recommends for turning a short idea into a full Krea 2 prompt, and its prompting guide says to use them “as a system prompt for LLM of your choice.” We did that for six short prompts, then rendered the short and the expanded version of each on Krea 2 Turbo, on four seeds.
Krea’s expansion prompt
This is docs/expansion.txt from Krea’s krea-2 repository, unchanged. Paste it into ChatGPT, Claude or a local model as the system prompt (or as your first message), then send your short idea. Copy the paragraph it writes into Krea 2.
You are an expert prompt engineer for text-to-image models. Your task is to expand the user's prompt into a highly effective image-generation prompt. Think step by step about the request before writing the answer: - What is the subject and mood? - What visual styles, mediums, and lighting options would fit? Consider two or three alternatives and pick the one that best serves the caption. - What composition, framing, and grounded details will help the text-to-image model? Then output a single expanded prompt paragraph. Follow these rules strictly: 1. **Faithfulness First:** Preserve all original subjects, actions, colors, and spatial relationships. Do not add new objects, props, characters, or animals unless the user clearly implies them. 2. **Practical T2I Structure:** Write a prompt that a text-to-image model can parse cleanly. Group subjects with their own attributes and actions. Use grounded phrasing for poses, interactions, and spatial layout. 3. **Style Planning Stays Internal:** Use your internal reasoning to choose style, medium, framing, and lighting. Do not emit planning tags or wrappers in the visible answer body. 4. **Text Rendering:** If the user requests visible text, quotes, labels, or typography, specify the exact text clearly and wrap requested words in quotes. 5. **Avoid Over-Specification:** Do not invent highly specific clothing, colors, materials, or scene details unless the input supports them. 6. **Structure:** Write one cohesive paragraph after the thinking block. No bullets, JSON, or markdown. 7. **Respect Existing Detail:** If the user's prompt is already detailed, lightly polish and finalize rather than heavily expanding — preserve their phrasing and direction. 8. **Respect the Human Form:** Treat depictions of people with dignity. Assume clothing covers genitals and intimate anatomy. 9. **Preserve User Medium:** When the user explicitly requests a medium (e.g. "photo of", "photograph of", "illustration of", "painting of", "sketch of", "3D render of"), honor it. Do not pivot to a different medium to avoid difficulty — match the user's stated intent.
Text by Krea.ai, from github.com/krea-ai/krea-2, published under the Apache License 2.0.
What we found
Krea’s expansion prompt is a way to choose the picture, not a quality button. Krea 2 followed the expanded paragraphs closely: most of the framing, light, lens and small details they described showed up on every seed. Whether that was an improvement depended on what the LLM chose.
- The seeds stop disagreeing. Every expanded prompt gave nearly the same frame on all four seeds. If you like to roll seeds for variety, keep the prompt short; if you want one specific picture, expand it. This matches the length test in our prompting guide.
- Vague phrases come out empty. “The usual ramen toppings” gave bare noodles, while the short prompt got a full bowl. Read the expansion before you use it and swap any hedge for real details.
- It decides things you didn’t say. It chose a pinkish-red for the neon sign and a side profile for the fox. Krea’s rule 5 tells the LLM not to over-specify, but it still has to choose a camera and a light.
- Short is sometimes as good. The lighthouse was already dramatic from eight words. On simple, familiar subjects Krea 2 Turbo’s own defaults are strong.
How we ran it
- The LLM: Claude Sonnet 5.5, one fresh session per prompt, with Krea’s text as its instructions and nothing added. We kept the paragraph it returned, unedited.
- The images: SirenRender’s free Quick Generate tile, Krea 2 Turbo at 8 steps and CFG 1, 960×1536, seeds 11, 22, 33 and 44. The short fox renders are the same four as in our prompting guide.
- Length: the expanded prompts came out 131 to 181 words, or 158 to 215 tokens with Krea 2’s Qwen3-VL tokenizer. That’s well under the 512 tokens Krea’s own code reads.
The six prompts, before and after
Fox in the snow
Short: a fox walking in the snow
A wildlife photograph of a red fox walking across an open expanse of fresh snow, caught mid-stride in side profile with one front paw lifted and the other paws pressing into the soft surface. Its head is held level and its gaze is directed ahead along its path. The fox's dense russet coat, white chest and full tail trailing low behind it stand out against the pale surroundings. Small puffs of loose snow kick up around its paws, and a line of paw prints trails behind it. Soft, diffused overcast winter light gives the scene cool, gentle tones and brings out fine detail in the fur and the fine snow crystals clinging to it. The shot is taken at the fox's eye level with a long telephoto lens at a wide aperture. The fox is sharply focused against a smooth, softly blurred snowy background. It sits slightly off-center, with open space in front of it, in a naturalistic, high-resolution documentary nature photography style.
The expansion turned a head-on close-up into the side-profile wildlife shot it describes: one paw lifted, paw prints trailing behind, soft overcast light. Both versions gave nearly the same picture on all four seeds, so the paragraph chose the shot rather than adding variety.
Lighthouse in a storm
Short: a lighthouse on a cliff during a storm
A dramatic cinematic photograph of a tall lighthouse standing on the edge of a rugged sea cliff during a violent storm. The lighthouse rises from the windswept clifftop in the right third of the frame, its lantern room glowing and sending a bright beam out through the rain and sea mist. Heavy, dark storm clouds churn across the sky, with a fork of lightning flashing in the distance and rim-lighting the cloud edges. Sheets of rain are driven sideways by the wind. Far below, large waves crash against the jagged, wet rock face of the cliff and burst into white spray. The camera is at a low angle with a wide lens, so the tower and cliff loom against the turbulent sky. The palette is cool blue-gray with deep shadows, set against the warm glow of the lamp, with sharp detail and moody atmospheric depth.
The short prompt already gave a dramatic storm, with spray, light beams and a keeper’s house. The expanded one put in everything it asked for on every seed: lightning, sideways rain, a lower camera and the tower in the right third. Moodier, but not clearly better. This is a case where the short prompt was enough.
Reader in a cafe
Short: a woman in her thirties reading a book in a cafe
A candid lifestyle photograph of a woman in her thirties sitting at a small cafe table, reading a book. Her head is tilted gently downward and her eyes are lowered to the open pages, with a calm, absorbed expression and natural skin texture. She holds the book open in both hands at a comfortable reading distance, her shoulders relaxed and her posture unhurried, as if lost in the story. The camera is at eye level in a medium three-quarter view, framing her from the head down to just below her hands, with the woman and the book in crisp focus. Soft natural daylight from a window to one side falls across her face and the pages, giving gentle highlights and smooth shadow transitions, while the cafe interior behind her melts into a warm, creamy blur of out-of-focus shapes and ambient light. Shot on a 50mm lens at a wide aperture with shallow depth of field, true-to-life color and fine detail, creating a quiet, intimate, unposed atmosphere.
Both read as a woman in her thirties reading. The short prompt changed the framing from seed to seed; the expanded one locked into the medium three-quarter shot, window light and blurred cafe it describes, the same frame on all four seeds, closer and more finished.
Bowl of ramen
Short: a bowl of ramen on a wooden counter
A photorealistic food photograph of a steaming bowl of ramen resting on a wooden counter, shot from a slightly elevated three-quarter angle so both the bowl and the counter surface are clearly visible. The deep bowl is filled with tangled, springy noodles in a glossy, savory broth that shimmers with tiny beads of oil, with the usual ramen toppings resting on the surface and soft wisps of steam curling upward. The wooden counter shows natural grain, subtle knots and gentle wear, and its warm tones lead the eye to the bowl, which sits near the center of the frame. Soft, warm window light falls from the side and slightly behind, catching the steam, adding gentle highlights on the broth and casting soft shadows across the wood. Shallow depth of field keeps the noodles and bowl rim in crisp focus while the background melts into a smooth, creamy blur of warm neutral tones. The mood is cozy, appetizing and inviting, with rich natural colors and fine detail in the noodle texture and wood grain, in the style of high-resolution editorial food photography.
Here the expansion made it worse. Krea 2’s own idea of ramen came with egg, sliced pork, nori and scallions on every seed. The expanded prompt only said “the usual ramen toppings”, and Krea drew bare noodles in broth on all four. If an LLM hedges like that, replace the vague phrase with the actual things.
Neon sign
Short: a neon sign that says "OPEN LATE"
A photograph of a glowing neon sign that reads exactly "OPEN LATE" in all capital letters, the two words set on a single line in clean, evenly spaced, easily legible block lettering formed from bent glass neon tubing. The sign is centered and fills most of the frame, mounted flat on a dark wall at night. The tubes burn a warm pinkish-red with a brighter, nearly white core, a soft halo and gentle bloom around every letter, and a faint wash of colored light spilling onto the wall around it. Shot straight on at eye level with a 50mm lens, sharp focus across the whole sign so each letter is crisp and fully readable, deep shadows and rich contrast, a quiet, inviting late-night mood in the style of cinematic night photography.
Both versions spelled OPEN LATE correctly on all four seeds. The short prompt gave the classic red-and-blue shop sign; the expanded one gave a moodier sign on a dark wall in the pinkish-red the LLM chose. It asked for the words on one line and got two lines every time.
Knight at dawn
Short: a knight on horseback at dawn
A cinematic photograph of a lone knight in full plate armor sitting upright on a strong horse, seen in side profile on a grassy rise and facing the sun as it clears the horizon ahead of them. Low golden-pink dawn light skims across the steel, picking out bright highlights along the helmet, shoulders, and gauntlets and along the horse's neck, mane, and flank. Long soft shadows stretch back across the dewy ground behind them. Thin mist pools in the valley below, and the sky melts from deep blue overhead into warm amber at the horizon. The horse stands calm and alert with its ears forward and a faint plume of breath in the cool air, and the knight keeps a steady, upright posture, gazing toward the rising sun. Wide framing with the pair placed slightly off-center, natural 35mm lens perspective, shallow depth of field with a softly hazy distant background, fine detail in the metal and the horse's coat, quiet and heroic mood.
The short prompt gave a game-cinematic knight with banners and a red horse cloth on every seed. The expansion’s side profile, low sun, mist and long shadows all appeared, and the result looks like a real photograph. Two flaws: the knight came out small for the horse on three seeds, and the horse’s “faint plume of breath” became large clouds of steam.
To write a long prompt by hand instead, the Krea 2 prompt builder assembles one from phrases we tested, and the prompting guide explains what each part does.