Grok Imagine Quality

Overview

grok-imagine-quality/image-to-image is xAI's high-fidelity version for image editing. It accepts 1 to 3 reference images and an editing instruction, and outputs finished images at 1K or 2K resolution. Compared with the standard version, it goes further in detail sharpness, texture and lighting, in-image text rendering, and composition stability, at the cost of slightly slower generation. It is suited to developers who need production-grade image quality and targeted modifications based on reference images, for delivery scenarios such as commercial visuals, social media assets, and layouts.

Key Highlights

• High-fidelity editing with more stable details and composition — Compared with the standard version, texture, lighting, and physical behavior are more natural and composition is more stable, so results can be used directly for delivery rather than only as drafts.
• Supports both 1K and 2K output — The same API lets you choose 1K for fast iteration or 2K for fine retouching, covering both creative trial and error and final production.
• Multiple reference images and controllable aspect ratio — Pass up to 3 reference images, and choose auto or 13 fixed aspect ratios; auto follows the first reference image.
• More accurate in-image text rendering — Text fidelity in layout-oriented images is improved, making it suitable for outputs containing text such as posters and social graphics.

Capabilities

Targeted editing with reference images: Pass 1 to 3 reference images along with a natural-language editing instruction to make targeted modifications to the image while preserving consistency of the subject and style.

Resolution and aspect ratio control: Supports 1K and 2K resolutions, and the aspect ratio can be set to auto or common portrait and landscape ratios; auto mode uses the aspect ratio of the first reference image, making it easy to adapt to different placements.

Multi-format output: Output formats include jpeg, png, and webp, directly matching the file type requirements of downstream processing pipelines.

Long-instruction editing: Editing instructions support up to 8000 characters, allowing description of fairly complex multi-element modification intents.

Named-entity fidelity: Accuracy is relatively high for named subjects such as brands, public figures, and specific locations, and objects can be named directly in the instruction.

Use Cases

• E-commerce and advertising teams modifying product images — Use product photos as references and adjust background, lighting, and composition according to a description to produce production-grade visual assets ready for launch.
• Designers creating posters and social media assets — Generate images with text layout based on existing layout reference images, and output 2K graphics for event posters or social media publishing.
• Creative teams iterating from sketches to fine retouching — Use 1K first to quickly validate direction, then output the final version at 2K, shortening the cycle from concept testing to delivery.
• Series images that need consistent style — Generate multiple images conditioned on the same reference image to keep style and subject consistent, for serialized content production.

Limitations and Selection Advice

This is an image editing pipeline and requires at least 1 publicly accessible jpeg, png, or webp reference image; SVG is not supported, so it is not suitable for pure text-to-image needs — for such scenarios, choose a text-to-image pipeline from the same series. Editing instructions support up to 8000 characters, and at most 3 images can be referenced per request. Quality is prioritized, so generation is slower than the standard version; high-resolution tasks or tasks requiring fine adjustment of reference images may need multiple attempts to reach the desired result. Only image output is supported and other modalities such as video are not covered.