Grok Imagine 2.0 Prompt Guide: Structure, Examples and Real Images

Write a Grok Imagine 2.0 prompt as plain sentences about one picture, in xAI's order: subject, details, style, light, color, framing. Put any words that must appear in the image in double quotes and say where they go. We ran 11 prompts on October 26, 2026: every quoted word came out right, and the model chose the image shape until we set aspect_ratio.

The Grok Imagine 2.0 prompt structure

xAI's Image Prompting Guide gives one order for every prompt in a series — subject, details and focal point, art direction, light, color, composition — and asks for four to six sentences of prose that describe one picture.

The last row, Text, is our addition.

• Subject first: xAI's guide says "A prompt that leads with mood, film stock, or a list of things to avoid produces a styled scene whose subject is a stand-in."
• Describe what you want, not what to avoid. "Brushed steel with fine horizontal grain" gives the model a material to draw, and a long avoid list is easy to ignore.
• Text: xAI's guide has no rule for words inside the image, and the announcement says the model plans typography and layout, so small text comes out sharp. We put every word in double quotes and said where it goes, and all 14 quoted lines came out right.
• Shape: the prompt does not set it. aspect_ratio defaults to auto, which lets the model pick the best ratio for the prompt, and our 11 unset runs came back portrait 832×1248, landscape 1248×832, square 1024×1024 and 1280×720. The API section shows how to lock the shape.
• Length and language: xAI publishes no maximum prompt length and no list of prompt languages. Ours were 30–48 words in one to three sentences, in English.

How we made the images

• Model grok-imagine-image-2.0 through OmniaKey's asynchronous media API on October 26, 2026, with operation generate and no parameters at all, so aspect_ratio stayed at xAI's default auto and the resolution at the default. Every task succeeded first time.
• Then two more runs of the poster prompt with aspect_ratio set to "16:9": one from our test harness and one from the Python script below. Both came back 1280×720 with the text spelled right.
• Output was JPEG, and the time from request to finished image ran from 15.6 s to 41.1 s. Most landed between 16 and 25 s; the icon took 41.1 s.
• We viewed every image at full size and read every word in the four text images against the prompt, letter by letter.
• OmniaKey sells API access to Grok Imagine 2.0, so we ran the prompts through it. Every price here is xAI's own.
• Sources: xAI's Image Prompting Guide, the image generation guide and the pricing page, checked October 26, 2026.

Each prompt ran once, so treat each image as one sample.

Eight prompts with real images

1. A poster with a title and one small line

This prompt asks for a big title, one small line and a flat screen-print look.

A minimalist jazz concert poster. Large bold title "BLUE HOUR" at the top and a smaller line "Live at Harbor Hall · Nov 14 · 8 PM" at the bottom. A single saxophone silhouette in deep blue on a cream background, lots of empty space, flat screen-print style.

Settings: generate, no aspect_ratio (auto), came back 832×1248 in 16.4 s.

What happened:

• "BLUE HOUR" sits in huge condensed dark-blue capitals across the top, the small line is at the bottom and spelled exactly, middle dots included.
• A dark-blue saxophone fills the middle of cream paper with a light print texture, and plenty of empty space stays empty.
• The saxophone shows its keys in light lines, so it reads as a drawing rather than a pure flat silhouette.

What to change:

• Nothing for the text; keep the small line in one pair of quotes so the dots stay with it.
• For a pure silhouette, try "a solid single-color saxophone shape with no inner details".

2. A product label with exact words

This prompt puts two short lines of text on a product photo.

Studio product photo of a glass bottle of cold brew coffee with a kraft paper label that reads "NORTH ROAST" and "Cold Brew · 12 oz". The bottle stands on wet dark slate with ice cubes around it, soft side light, dark background, sharp focus on the label.

Settings: generate, no aspect_ratio (auto), came back 832×1248 in 17.1 s.

What happened:

• A dark glass bottle of coffee covered in water drops stands on wet black slate with ice cubes around it, against a black background.
• The kraft label reads "NORTH ROAST" in bold capitals and "Cold Brew · 12 oz" under a thin line, both exact.
• The model also drew a small mountain outline above the name, which the prompt did not ask for.

What to change:

• If the label must carry only your words, try "the label shows only these two lines of text, no logo".

3. A recipe card: several lines and icons

This prompt asks for a title, three numbered steps and one icon per step.

A clean recipe card titled "3-Step Pancakes" with three numbered steps: "1. Mix flour, milk and eggs", "2. Rest the batter 10 minutes", "3. Cook on a hot pan". A small line icon next to each step, cream background, simple sans-serif font.

Settings: generate, no aspect_ratio (auto), came back 832×1248 in 15.6 s.

What happened:

• The title "3-Step Pancakes" sits at the top over a short tan rule, and three rows are split by thin lines.
• Each step's words are exact and wrap over two lines, with a line icon on the right: a whisk, a clock, a frying pan.
• The numbers sit in circles ("1", "2", "3") instead of "1.", "2.", "3.".

What to change:

• If the "1." style matters, try "numbers written as 1., 2., 3. before each step, no circles".

4. A chalkboard menu with prices

This prompt tests five short lines of handwriting with prices.

A cafe chalkboard menu with the heading "TODAY" and five lines of small handwritten text: "Flat white 4.50", "Oat latte 5.00", "Matcha 5.50", "Almond croissant 4.00", "Banana bread 3.50". Wooden frame, warm cafe light.

Settings: generate, no aspect_ratio (auto), came back 1248×832 in 29.8 s.

What happened:

• A blackboard in a rustic wooden frame hangs on a warm plaster wall, with a potted plant at the left edge.
• "TODAY" is in large chalk capitals with an underline and a small leaf doodle.
• The five items are in neat chalk handwriting, the prices line up in a column on the right, and every item and price is exact.

What to change:

• Keep each line short and give the exact words; five short lines all came out right here.

5. A street photo: lens and light

This prompt uses camera words to ask for a photographic look.

Candid street photograph of an elderly man feeding pigeons in a stone plaza at golden hour. Shot on a 35mm lens, shallow depth of field, natural skin texture, warm low sun, slight film grain.

Settings: generate, no aspect_ratio (auto), came back 1248×832 in 19.8 s.

What happened:

• An old man in a flat cap and a long dark coat crouches on stone paving and holds out an open hand to the pigeons around him.
• The light is warm and low, with a blurred plaza, people and a large stone wall behind him.
• His face shows wrinkles and skin texture and the hand looks natural; the film grain stays light.

What to change:

• None needed. xAI's guide says to state texture once and ask for light that keeps the subject readable, and both held here.

6. Food from above: name every ingredient

This prompt lists the toppings and gives the light a direction.

Overhead food photo of a bowl of shoyu ramen with a soft-boiled egg, chashu pork, nori and sliced scallions on a dark wooden table. Chopsticks rest on the bowl, light steam rises, natural window light from the left.

Settings: generate, no aspect_ratio (auto), came back 1024×1024 in 20.7 s.

What happened:

• A straight-down view of a stoneware bowl on dark weathered wood: half an egg with an orange jammy yolk, two slices of rolled chashu, two sheets of nori, a heap of sliced scallions and noodles in brown broth.
• Wooden chopsticks lie across the right side of the bowl, with faint wisps of steam at the top.
• The model added a pile of dark shredded mushroom, which the prompt did not list.

What to change:

• List everything that should be in the bowl, then try "nothing else on top" if the toppings have to match exactly.

7. An interior: each object with its material

This prompt names every object with its material and color.

Interior design photo of a small Scandinavian living room: light oak floor, a white sofa with a mustard throw, a round oak coffee table, a large window with soft morning light, a tall fiddle-leaf fig in a white pot. Calm, airy, eye-level view.

Settings: generate, no aspect_ratio (auto), came back 1248×832 in 16.3 s.

What happened:

• Every named item is there: light oak floor, white sofa with a mustard throw over one arm, round oak table on a woven rug, the big window on the left with trees outside, a tall fiddle-leaf fig in a white pot on the right.
• Framed art on the wall, an eye-level view, bright and soft.

What to change:

• None needed. A prompt that names each object, its material and its color got all of them; keep that pattern for other rooms.

8. A watercolor illustration

This prompt names a medium and a paper texture.

Children's book illustration in soft watercolor: a small fox reading a book under a giant red mushroom in a forest, fireflies glowing around it, gentle pastel colors, visible white paper texture.

Settings: generate, no aspect_ratio (auto), came back 832×1248 in 24.9 s.

What happened:

• A soft watercolor of an orange fox under a big red-capped mushroom, holding an open blue book, with glowing fireflies around it, ferns, small flowers and little mushrooms.
• White paper texture shows at the edges.
• The fox's eyes are closed, so it looks more asleep than reading.

What to change:

• Say the expression you need: try "eyes open, looking down at the page".

What goes wrong and how to fix it

A vague prompt lets the model choose everything

Every word you leave out, the model fills in, and it can fill it with a real brand.

a cool car

A 1960s cherry-red convertible parked on a coastal road at sunset, low angle from the front three-quarter view, chrome bumper catching the light, ocean and cliffs behind it, photorealistic, warm golden light.

• "a cool car" came back as a dark grey modern supercar on wet asphalt at dusk, with mountains, pine trees and a cloudy pink sky behind it. It carries a real car maker's badges on the hood and the wheels.
• The specific prompt put everything in the picture: the cherry-red convertible, the low front three-quarter angle, the chrome grille and bumper in the sun, the sea and cliffs behind. The model added a number plate reading "1962 CLK".

Simple icons can look like famous logos

The icon prompt asked for plain shapes and got them, close to a mark you already know.

Flat vector app icon: a white paper plane on a rounded square with a blue-to-teal gradient. Simple geometric shapes, no text, centered, clean edges, plain white background around the icon.

The result is exactly what was asked — a white paper plane, slightly off-center, on a rounded blue-to-teal square, no text — and it looks very close to a well-known messaging app's logo. Simple, common icon ideas land near famous marks, so add one detail that is yours and compare the result with existing logos before you use it.

These are the misses we saw, with the fix for each.

Run these prompts through the API

OmniaKey's media API takes three steps: create a task, poll the same task until it finishes, then download the asset.

pip install requests
export OMNIAKEY_API_KEY="your-key"

Here is the script we ran, unchanged:

import os
import time
import uuid

import requests

API = "https://api.omniakey.com/v1/media/tasks"
HEADERS = {"Authorization": f"Bearer {os.environ['OMNIAKEY_API_KEY']}"}

PROMPT = (
    'A minimalist jazz concert poster. Large bold title "BLUE HOUR" at the top and a smaller '
    'line "Live at Harbor Hall · Nov 14 · 8 PM" at the bottom. A single saxophone silhouette '
    "in deep blue on a cream background, lots of empty space, flat screen-print style."
)

# 1. Create the task. Reuse the same Idempotency-Key only when retrying this exact request.
resp = requests.post(
    API,
    headers={**HEADERS, "Idempotency-Key": str(uuid.uuid4())},
    json={
        "model": "grok-imagine-image-2.0",
        "type": "image",
        "operation": "generate",
        "input": {"prompt": PROMPT},
        "parameters": {"aspect_ratio": "16:9"},  # leave it out and the model picks the shape
    },
    timeout=30,
)
resp.raise_for_status()
task = resp.json()
print("task:", task["id"], task["status"])

# 2. Poll the same task until it finishes. Slow is not failed: keep polling this id.
while task["status"] in ("queued", "processing"):
    time.sleep(4)
    task = requests.get(f"{API}/{task['id']}", headers=HEADERS, timeout=30).json()

if task["status"] != "succeeded":
    raise SystemExit(f"task ended {task['status']}: {task.get('error')}")

# 3. Download the image now: the signed URL lasts about an hour.
asset = task["assets"][0]
with open("blue-hour.jpg", "wb") as f:
    f.write(requests.get(asset["url"], timeout=60).content)
print("saved blue-hour.jpg", asset["mime_type"], asset["byte_size"], "bytes")

Its output on October 26, 2026:

task: task_ETAcsodTDIPQkhTVyykyRiRdBpYDOgj7 queued
saved blue-hour.jpg image/jpeg 205099 bytes

From request to saved file took about 16 seconds.

• aspect_ratio: xAI's API accepts 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 2:1, 1:2, 19.5:9, 9:19.5, 20:9, 9:20, 21:9, 5:2 and auto. xAI suggests locking one ratio across a series and 3:2 for most photographs. Our two 16:9 runs both came back 1280×720, and the layout was rebuilt for the wide frame rather than cropped: title across the top, saxophone in the middle, the small line at the bottom.
• On OmniaKey, Grok Imagine is generation only and n must be 1. xAI's own API also edits images and returns up to 10 per request.
• xAI's API also has quality and resolution settings, which we did not test through OmniaKey.
• Download the file when the task finishes: the signed link lasts about an hour and the task record 24 hours.
• Price: xAI lists Grok Imagine 2.0 at $0.04 per image, flat whatever the prompt length (checked October 26, 2026). The older grok-imagine-image-quality is $0.05 and grok-imagine-image is $0.02 per image, and our Grok Imagine API test of the three image models runs the same four prompts through all three. On OmniaKey a failed task is refunded automatically. For OmniaKey's current rate, see the Grok Imagine 2.0 model page and the image API docs.

Common questions

How long can a Grok Imagine 2.0 prompt be?

xAI publishes no maximum length for an image prompt. Its Image Prompting Guide asks for four to six sentences for a single still, and ours were 30–48 words in one to three sentences. All eight recipes above still landed close to the prompt.

How do you get text right in Grok Imagine 2.0?

Put each line of text in double quotes and say where it sits in the frame. Our four text prompts rendered all 14 quoted lines correctly, middle dots and prices included, and xAI's announcement says small text now comes out sharp. xAI's guide has no rule for text, so the quotes are our method, not xAI's.

What size are Grok Imagine 2.0 images?

Left on auto, our 11 runs came back at 832×1248, 1248×832, 1024×1024 and 1280×720. Setting aspect_ratio to "16:9" gave 1280×720 in both runs. xAI's API also has resolution, with 1k as the default and 2k as an option.

Can Grok Imagine 2.0 edit an image?

Yes through xAI, which takes an edit at POST /v1/images/edits with up to 5 source images in one request. Its guide says to lead with the one to three changes and keep the rest. On OmniaKey, Grok Imagine is generation only, so an edit request is rejected.

Does Grok Imagine 2.0 rewrite my prompt?

xAI's API bills prompt text for a prompt-rewriting (upsampler) model, and its response schema has no field for a rewritten prompt. What the upsampler changes is not published, so read the image against your prompt instead of guessing.

How much does Grok Imagine 2.0 cost?

xAI lists $0.04 per image for grok-imagine-image-2.0, checked October 26, 2026, and generation costs the same whatever the prompt length. An edit is billed for the input image and the output image. For OmniaKey's current rate, see the Grok Imagine 2.0 model page.

Update log

• October 27, 2026: first version, with prompts run on October 26, 2026.