Grok Imagine Image 2.0 Preview

Introduction

Grok Imagine Image 2.0 is xAI's image model built for images you can actually ship, not just pretty renders. It reads a prompt the way a photographer reads a brief, holds the subject, the framing, and the light together, and keeps the small real-world details right that most models fumble. It ranks second on the public text-to-image and image-edit leaderboards, and the gap it closes is factual, production-ready detail.

A cinematic full-length fashion editorial photograph of a model in a tailored oversized camel wool coat over a cream ribbed knit and wide-leg trousers, standing on a rain-slicked city street at blue hour, warm shop lights glowing softly behind her, shallow depth of field, directional light catching the weave of the wool, muted sophisticated colour grade, medium-format sharpness, natural skin texture, editorial magazine quality.

This guide covers the request shape, how to structure a prompt, directing the camera and light, the fixed 1K and 2K size pairs, and leaning on the model's factual accuracy. Rendering readable text in the image is its own strength with its own text guide, and editing an existing image is covered in the editing guide.

The request

A text-to-image call is an imageInference task. Only positivePrompt is required, and for text-to-image you also pass an explicit width and height.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'xai:grok-imagine@image-2.0',
  positivePrompt: 'A cinematic full-length fashion editorial photograph of a model in an oversized camel wool coat and cream knit on a rain-slicked city street at blue hour, warm shop lights behind her, shallow depth of field, medium-format sharpness, editorial magazine quality.',
  width: 1776,
  height: 2368
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "xai:grok-imagine@image-2.0",
            "positivePrompt": "A cinematic full-length fashion editorial photograph of a model in an oversized camel wool coat and cream knit on a rain-slicked city street at blue hour, warm shop lights behind her, shallow depth of field, medium-format sharpness, editorial magazine quality.",
            "width": 1776,
            "height": 2368
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "imageInference",
      "taskUUID": "b1f2a3c4-d5e6-4789-a012-3456789abcde",
      "model": "xai:grok-imagine@image-2.0",
      "positivePrompt": "A cinematic full-length fashion editorial photograph of a model in an oversized camel wool coat and cream knit on a rain-slicked city street at blue hour, warm shop lights behind her, shallow depth of field, medium-format sharpness, editorial magazine quality.",
      "width": 1776,
      "height": 2368
    }
  ]'

runware run xai:grok-imagine@image-2.0 \
  positivePrompt="A cinematic full-length fashion editorial photograph of a model in an oversized camel wool coat and cream knit on a rain-slicked city street at blue hour, warm shop lights behind her, shallow depth of field, medium-format sharpness, editorial magazine quality." \
  width=1776 \
  height=2368

{
  "taskType": "imageInference",
  "taskUUID": "b1f2a3c4-d5e6-4789-a012-3456789abcde",
  "model": "xai:grok-imagine@image-2.0",
  "positivePrompt": "A cinematic full-length fashion editorial photograph of a model in an oversized camel wool coat and cream knit on a rain-slicked city street at blue hour, warm shop lights behind her, shallow depth of field, medium-format sharpness, editorial magazine quality.",
  "width": 1776,
  "height": 2368
}

[
  {
    "taskType": "imageInference",
    "taskUUID": "b1f2a3c4-d5e6-4789-a012-3456789abcde",
    "imageUUID": "7c1b2d3a-4e5f-6789-abcd-ef0123456789",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/7c1b2d3a-4e5f-6789-abcd-ef0123456789.jpg"
  }
]

Four things shape a text-to-image request:

• positivePrompt is required. There is no character ceiling to worry about in practice, but detail past a point competes with itself, so keep every clause carrying weight.
• width and height are a fixed pair, not free numbers. Each pair is one of 13 aspect ratios at either the 1K or 2K tier, listed under Sizes and aspect ratios. Pass a pair that isn't on the list and the request is rejected.
• resolution (1K or 2K) is a shorthand that only applies when editing an image. For text-to-image it isn't accepted, and it can't be combined with width and height. The editing guide uses it.
• settings.quality currently exposes a single medium level, so there is nothing to tune there yet.

Structuring the prompt

The model rewards a prompt that separates the subject, the setting, the composition, the light, and the finish. Each becomes a directive it tries to satisfy, so naming all five keeps the decisions with you instead of the model's defaults.

An overhead editorial food photograph of a weekend brunch spread on a pale oak table: a short stack of golden buttermilk pancakes with a pat of melting butter and a thin drizzle of maple syrup, a small ceramic bowl of fresh raspberries and blueberries, a glass of fresh orange juice, and a loosely folded linen napkin. Bright soft natural window light from the left, fresh airy colour grade, shallow depth of field, crisp realistic detail, clean and appetizing.

Each layer earns its place. The subject fixes what the shot is about, the setting grounds it in a real place, and the composition and light are what separate a snapshot from a shot that looks directed. The finish clause sets the register, editorial photography versus product render versus flat illustration, and skipping it hands the look to the model.

Directing the shot

Grok Imagine reads photographic language directly. Name the framing, the lens, and the light in the terms a photographer uses and it applies them rather than guessing.

A wide architectural interior photograph of a modern open-plan living room at golden hour, low warm sunlight raking across a pale oak floor and a linen sofa, floor-to-ceiling windows framing a soft city view, a large abstract canvas on the far wall, styled with a few plants and ceramics. Symmetrical eye-level composition, wide-angle lens with straight verticals, warm natural light, realistic materials and reflections, high-end real-estate and interior editorial quality.

The terms that land: framing and shot size (wide shot, close-up, overhead flat-lay, three-quarter portrait), lens behaviour (wide-angle with straight verticals, telephoto compression, shallow depth of field), camera height and angle (eye-level, low angle, top-down), and the light (golden hour, soft window light, hard studio key, backlit). Pair the framing with the light and the model composes to both. Leave them out and it defaults to a centered, evenly-lit middle distance.

Sizes and aspect ratios

Text-to-image runs on fixed size pairs, one width and height per aspect ratio, at two tiers. 1K is the fast everyday tier, and 2K roughly doubles each edge for print, large crops, and fine detail. Both cover the same 13 aspect ratios:

• Square and near-square: 1:1, 4:3, 3:4.
• Photo and screen: 3:2, 2:3, 16:9, 9:16.
• Wide and tall extremes: 2:1, 1:2, 19.5:9, 9:19.5, 20:9, 9:20.

The aspect changes what the model puts in frame, not just the crop. The same mountain scene fills a wide frame with the valley, centers the figure in a square, and stacks the ridge vertically in a tall one:

A cinematic landscape photograph of a lone hiker in a red shell jacket standing on a rocky mountain ridge at sunrise, a vast valley of layered blue peaks and low drifting mist below, warm golden light breaking across the range, crisp clear alpine air, natural realistic detail, a strong sense of scale and stillness.

A cinematic landscape photograph of a lone hiker in a red shell jacket standing on a rocky mountain ridge at sunrise, a vast valley of layered blue peaks and low drifting mist below, warm golden light breaking across the range, crisp clear alpine air, natural realistic detail, a strong sense of scale and stillness.

A cinematic landscape photograph of a lone hiker in a red shell jacket standing on a rocky mountain ridge at sunrise, a vast valley of layered blue peaks and low drifting mist below, warm golden light breaking across the range, crisp clear alpine air, natural realistic detail, a strong sense of scale and stillness.

Pick the aspect for where the image lands, a 16:9 hero, a 1:1 social post, a 9:16 story, a 2:1 banner, and let the model recompose for it rather than cropping a single render by hand.

Leaning on factuality

The model's headline improvement is factual accuracy: believable proportions, plausible scenes, and the small details a client would catch. It holds up on the shots that usually expose a model, dense crowds, real architecture, legible signage, correct reflections, and everyday objects that have to look right rather than merely plausible.

A photoreal street-level photograph of a busy pedestrian crossing in a modern city at rush hour: commuters mid-stride carrying bags and coffees, yellow taxis and a red double-decker bus waiting at the light, glass office towers reflecting a bright overcast sky, storefronts and legible street signage, wet pavement with soft reflections. Accurate architecture, believable crowd detail and correct proportions, documentary editorial style, natural colour, sharp from front to back.

The practical payoff is fewer regenerations on the shots that matter. When a scene has a lot going on, name the elements that have to be right, the vehicles, the storefronts, the number of people, the reflections, and the model keeps them consistent rather than inventing a vaguer scene. For the same reason it is a strong base for real-work deliverables where a wrong detail is a reshoot.

Tips

1. Name all five layers. Subject, setting, composition, light, and finish. Dropping any one hands that decision to the model's defaults, and the finish clause is the one most often forgotten.

2. Direct the shot in photographic terms. Framing, lens, camera height, and light land as instructions. "Wide-angle, eye-level, golden hour" beats leaving the shot to a centered, evenly-lit default.

3. Pass a valid size pair. width and height must be one of the listed 1K or 2K pairs. Iterate at 1K, then re-render the keeper at 2K for print and large crops.

4. Match the aspect to the destination. The model recomposes for each aspect, so ask for the 9:16 or 2:1 you actually need instead of hand-cropping a 1:1.

5. Call out the details that must be right. On busy or real-world scenes, list the elements a client would check, and the model holds them steady instead of drifting to a vaguer version.

6. Keep resolution for editing. In text-to-image it is not accepted. Use width and height here, and reach for resolution only when editing an image.