Splitting a character image into 18 layer-ready parts with GPT Image 2.5
A long ChatGPT prompt for cutting one illustration into transparent parts that stack back into the original, and why it reads like a specification.
1mm Module (@1mm_module), who writes in Japanese, published a long prompt that asks ChatGPT with GPT Image 2.5 to break one character illustration into 18 transparent PNGs that stack back into the original, the kind of part set used to rig a 2D avatar. The author says they tested up to 27 parts, including left and right splits, but compositing accuracy dropped sharply past about 20, so the published version stops at 18.
The author states that results are not guaranteed, and adds that the quality of the output changes visibly with the reasoning effort setting, enough that the task might work as a benchmark. Neither claim was tested here. The prompt is in Japanese and is linked below; this note summarizes its structure instead of translating it.
Pin the original and the canvas
The prompt treats the attached image as the only source of truth. Before any work, the model records the file name, the real width and height, the color mode, whether there is alpha, and a SHA-256 hash. Every part, mask, and check image must use exactly that canvas, with the origin at the top left. It explicitly forbids centering parts automatically, trimming them to the opaque area, or scaling each one to fill the frame, because any of those breaks the stack.
Decide ownership before generating
The 18 parts run from headwear, eye whites, brows, lashes, irises, nose, mouth, and face through front and back hair, neck, top, shoulders, arms, hands, lower wear, legs, and footwear. The model must first map each part’s visible area, hidden area, stacking order, and anchor points, and assign pixels by outline and continuity rather than by color. An outline belongs to the object in front, so no edge is drawn twice.
Keep what is visible, paint what is hidden
Each part is treated as three regions. The visible interior keeps the original pixels. The blended edge gets a smooth alpha instead of a hard threshold. The hidden region must be painted in: skin under the bangs, eye white behind the iris, hair behind the shoulders. Leaving a hole or smearing nearby color does not count. The eyes get their own checks against masks traced from the original.
Verify like a pipeline
The prompt spells out the compositing math, allows the original’s alpha to be restored only once all 18 parts exist, and bans tricks that fake a match, such as pasting the original back over the result. It works in batches of up to ten parts, reloads every saved part from disk before each delivery, composites them over white, black, and gray, and nudges each part by a few pixels and degrees to expose holes. Results are reported as separate pass or fail items, and anything not checked is marked as unchecked.
Read it before you paste it
The published text still says 27 parts in one heading and two instructions, plans the last batch as seven images instead of eight, and names the face part base_face in one table. Fix those before using it. The habits it teaches carry over to any image edit that has to be exact: pin the source, name the shortcuts you forbid, and require checks on real files rather than a description of them.
A condensed brief to adapt
A short version in our own words. It does not replace the full specification.
Treat the attached image as the only original. Record its file name, exact width and height, and whether it has alpha. Split the character into these transparent PNG layers: [list of parts]. Every layer must use the original’s exact canvas and position: no cropping, centering, or rescaling. Assign pixels by outline and continuity, not by color, and give each outline to the object in front. Paint in the areas hidden behind other parts so that no holes appear when a layer moves slightly. Do not paste the original back onto the result to fake a match. Before delivering, reload every saved layer, stack them in order over white, black, and gray, compare the result with the original, and list what passed, what failed, and what you did not check.