Generate Four Grok Branches Before Editing One

Batch generation is useful when it creates comparable branches, not a pile of unrelated pictures. Hold the brief constant, make four Grok images, and choose the composition that needs the least structural repair.

A single output encourages premature commitment. The first credible frame starts collecting edits, and after three rounds the time already invested can feel like evidence that it deserves another. Grok Imagine's image-generation API supports multiple outputs from a request. Asking for four is a practical editorial choice: enough variation to expose alternate compositions, small enough to compare carefully.

Keep the generation brief identical across the branch. Same aspect ratio, subject, camera, light, object count, exclusions, and copy field. If you also change the wording for every candidate, you are comparing prompts and images at the same time. The goal is to see which visual solution Grok produces from one set of constraints, then select rather than negotiate with the first result.

xAI documents batch image generation and separate output assets for Grok Imagine. Four branches and the five-part scorecard are Luxe Prompting's editorial method; four is a recommended comparison size here, not a stated platform maximum.

xAI Docs — Grok Imagine image generation

xAI Docs — Imagine file outputs

xAI Docs — Grok Imagine image editing

Use five zero-to-two scores. Hierarchy: is the intended hero first? Geometry: are shapes and contacts credible? Crop: are important edges safe? Copy space: is the assigned field genuinely usable? Edit distance: how many structural repairs stand between this image and delivery? A perfect score is ten. Write the numbers before discussing mood, polish, or personal favorite.

The scorecard prevents one dramatic highlight from hiding a broken handle or crowded headline area. It also reveals ties. If two frames both score eight, choose the one whose failure is local. A wrong reflection can be a bounded edit; a wrong camera angle changes the whole image. Select for the best structural future, not the strongest first impression.

Save the winning output as its own reference before editing. xAI documents file-based image workflows, and each generated output can be handled separately. In the edit request, name the source image, the exact failing region, the replacement behavior, and the passing relationships that must remain. Keep the command smaller than the original generation brief.

For example: “On the selected image, repair only the watch crystal so it has one clean window reflection and no cloudy center. Preserve the 3:2 crop, camera angle, hand position, watch geometry, dial markings, background gradient, and right-side copy space.” After the edit, rescore all five criteria. A local repair that damages geometry or hierarchy is not a successful edit.

Before selection: Were all four generated from the same brief? Are they shown at equal size? Did you score hierarchy before detail? Did you inspect hands, joins, reflections, repeated objects, and crop boundaries? Is the chosen image best by the written score rather than merely the first one you opened? Record the winner and its exact failure.

After editing: Did only the named region change? Are camera, silhouette, object count, type-safe space, and light map intact? Did a repaired surface create a new contact error nearby? If the edit loses a passing relationship, return to the original winning output and tighten the boundary. Do not keep repairing a repair whose foundation has drifted.

Use this unchanged generation brief for all four candidates. Apply the repair paragraph only after the scorecard names a winner.

Generate four 3:2 editorial product images from this same brief: one fictional brushed-steel wristwatch resting across a folded charcoal linen cloth on a pale stone surface; camera 20 degrees above table level; watch occupies the left 52% of frame; dial faces camera and remains circular; right 32% is quiet low-contrast copy space; one broad softbox reflection from upper left; restrained shadow beneath the case; no wrist, box, logo, lettering, extra watch, jewelry, flowers, smoke, splash, or border. PRIORITIZE: clean silhouette, credible strap-to-case joins, readable dial geometry, safe crop, and empty right field. WINNER EDIT ONLY: repair only [named region and behavior]. Preserve camera, watch geometry, hand positions, cloth folds, object count, background, light direction, crop, and copy space.

The load-bearing line: “Preserve camera, watch geometry, hand positions, cloth folds, object count, background, light direction, crop, and copy space.” The first paragraph creates comparable branches. The final line makes the chosen edit accountable to every relationship that already passed.

Take one Grok image brief and request four branches without changing the wording. Score them, then change only the failing region variable on the winner.

CHANGE ONLY: FAILING REGION VARIABLE