SheetSlicer

ChatGPT Sprite Sheet Uneven Grid: Diagnose Before Cropping

Check 3 things before cropping an uneven AI sheet: whether cells share a measurable grid, whether complete poses are separated, and whether any artwork is already missing.

Updated

Why a requested grid is not a measured grid

A prompt can describe four columns and two rows without the resulting image honoring those boundaries. Canvas dimensions may divide neatly while the drawings still drift across the implied cells. Separate the layout you requested from the pixels you actually received. The important question is not whether the prompt says “eight frames,” but whether eight complete poses can be extracted.

This guide is a decision and crop workflow, not another prompt collection or a promise about a particular model. The broader ChatGPT sheet repair guide covers cleanup settings. Here, first decide whether a crop-based tool can solve the problem at all. No automatic crop can recreate a missing hand or determine ownership of genuinely overlapping pixels.

Step by step: classify before cutting

  1. Download the original image and preserve it. Note its actual dimensions and count the intended poses. A screenshot may include a display checkerboard or may have been rescaled, making otherwise reasonable grid calculations misleading. Do not start by shrinking the art to match an assumed cell size.
  2. Examine the first, middle, and last apparent cell boundaries. Look for equal spacing, consistent outside margins, and intact limbs. If every complete pose fits a regular rectangle, use Manual grid in the slicer. If each successive crop shifts farther off target, recheck gaps and offset before trying a different mode.
  3. If poses are unevenly spaced but separated by clear background, try Auto full poses. It follows connected foreground regions rather than enforcing your requested grid. Compare the outlines with the intended characters; orange overhang indicators are a reason to inspect, not an automatic proof of a good detection.
  4. If a detached particle, weapon, or shadow becomes its own detection, decide whether a known grid can keep it with the correct pose. Do not delete it merely to make the count equal eight. If no nonoverlapping rectangle contains the intended frame, assemble that frame in another art editor.
  5. If neighboring poses touch, inspect the actual contact point. A blank separating line in an editor may solve an accidental bridge, but shared anatomy or overlapping silhouettes require manual repair or regeneration. Repeatedly switching grid sizes is not a recovery strategy for missing source pixels.
  6. After crops are correct, address background and alignment. Choose a low-tolerance solid key only when appropriate. Trim and align grounded poses if that removes accidental drift, but preserve intentional vertical motion for jumps and effects. Crop correctness and motion correctness are separate checks.
  7. Reorder the filmstrip and preview slowly with sprite sheet to GIF. Compare the largest and smallest poses and the final-to-first transition. Stop and edit the source if scale, costume, viewpoint, or anatomy changes rather than trying to hide those differences with a faster FPS.

A practical decision example

Suppose an image is 1024 × 512 and appears to contain eight poses. A four-by-two division gives nominal 256 × 256 cells; that is a calculated example, not a generator guarantee. If a scarf crosses a nominal boundary but stops before the next pose, connected detection may preserve it. If the scarf merges into the next character, crop settings alone cannot identify the missing separation.

Reference checked November 9, 2026: Aseprite's sheet import documentation identifies offsets, frame dimensions, padding, and layout order as distinct import inputs. That reinforces the geometric distinction used here; it does not verify a ChatGPT output-size contract or imply that an art editor understands AI anatomy automatically.

Common pitfalls

Do not treat the detected frame count as the acceptance test. Eight detections can include a loose sword and omit a merged pair of characters. Check complete content inside every frame. Likewise, a perfectly uniform output atlas can contain consistently cropped-off feet; repacking does not prove the source was sliced correctly.

Do not use bottom alignment to normalize intentionally different actions. An attack effect below the feet can change the visible lower bound, and a jump should not necessarily stay on the ground. Compare with the original pose sequence before accepting a cleaner-looking but mechanically incorrect animation.

Related tools and next action

Use the background remover for an actual solid background, not a painted checkerboard. Use the packer after manually repairing separate frames. If regeneration is necessary, the prompt guide helps describe a clearer frame contract without claiming that any model will obey it perfectly.

Frequently asked questions

The dimensions divide evenly. Why do crops still cut off limbs?

Divisible canvas dimensions do not prove that the drawings follow the grid. Inspect actual pose bounds and try full-pose detection when the poses are separated.

Is eight detected frames proof that an eight-pose sheet is correct?

No. A detached object can count as a frame while touching poses merge. Review the contents of each detection.

When should I stop adjusting the grid?

Stop when poses overlap, source pixels are missing, or character identity changes. Those need art editing or regeneration rather than another crop setting.

Free sprite tools

More guides