গ্ৰক ইমাজিন 2.0

xAI's current recommended image model, a second-generation Grok Imagine that improves prompt adherence and detail over the standard tier. It generates and edits images from text and reference images across the same aspect ratios.

xAI states the goal for this model plainly: images you can use in real work. The lab's claim is about composition rather than raw fidelity: type placed and a layout arranged with intent, which is what keeps a crowded multi-panel image coherent and its small lettering legible. It also says the model holds on to whatever you hand it, both when generating and when you come back to edit. Training covered photographic, design and illustrative registers alike, and xAI presents editing as something the model was built to do rather than a mode bolted on later.

Most of what changed is about iteration. xAI designed around the assumption that nobody ships the first render, so the emphasis is on narrow, repeatable changes: a wand that touches only the area you indicate, a segmentation tool for picking out a precise region, and background removal that returns the subject on transparency ready to drop into other work. Through the API the same model accepts up to five source images in a single request, which removes the manual compositing step from work that used to need it.

Framing is a parameter here rather than a crop. Pick an aspect ratio and the model fills the frame; the range runs from tall portrait through square to very wide cinematic formats, plus an automatic setting that follows the first reference image you supply. Output comes back at 1K or 2K, as PNG or JPEG.

xAI also shipped templates - packaged starting points for photo editing, product shots, headshots, icons and game assets, each with the workflow already configured - and demonstrated a character, her locations and the props she carries generated separately while holding a single visual style from image to image. That style-holding across separate generations is what makes the reference-image path useful for brand and character work rather than for one-off pictures.

The limits are worth knowing before you build around it. There is no mask: every edit is described in words and located by the model, so an instruction that is ambiguous about which region you mean gets resolved by the model rather than by you. There is no seed, so the same prompt does not reproduce the same image twice. Requests return one image at a time, and reference images are capped at five, 20 MB each, PNG or JPEG only.