In six numeric-prompt trials, opposite proportion instructions produced figures within the measurement margin of error. For this style and setup, silhouette language was more useful than numeric ratios for changing the character’s proportions.
Six Ways of Asking
The subject was kept deliberately dull so that nothing but proportion was in play: a circus strongman in a striped leotard on a small round stage, made on the AI image generator at 16:9, plain warm backdrop, soft even light. One paragraph of scene description stayed byte-for-byte identical in all six. Only the proportion sentence at the end changed.
What the Numbers Did
The six numeric runs and, at the right, the one that worked. Height in head-heights is measured between the marked lines.
Runs 1 to 6 measured 4.4, 4.7, 4.5, 4.3, 4.5, and 5.0 heads tall. Because the estimated measurement error is about ten percent, these results do not show a reliable proportion change for this style and test setup.
The landmarks here were read off a coordinate grid laid over each picture and then drawn back onto it and checked before any number was used, which puts the reading error at roughly twenty pixels per landmark, or about ten percent on a head-count. The whole spread across the six runs sits inside that. The honest summary is that none of the six instructions moved the figure by an amount this measurement can detect.
The Proportions Were Already Chosen
Run 1 is the interesting one. It asked for normal human proportions and returned a figure 4.4 heads tall with shoulders far wider than its hips — a stocky cartoon build, not a human one. An adult human runs to roughly seven or eight heads. The style words in the shared paragraph — a 3D cartoon character, a stylized toy-like 3D animation look — had already decided the build before the proportion sentence was reached.
That is why run 2 had nothing to do. It asked for shoulders twice the width of the hips on a figure whose shoulders were already wider than that. And it is why run 4 had nothing to add: the character was already a caricature, so asking for caricature changed little.
The Sentence That Moved It
Run 7 dropped every number and described the silhouette in ordinary words instead.
Run 7. His head is enormous and sits straight on his shoulders with no neck at all, so that his face fills the upper part of the figure. His arms are short and stubby, his legs are short, and his boots are tiny.
That worked immediately, and it took three more runs to make it usable.
Runs 7 to 10. The proportion arrived on the first of them; the rest is about framing and about which direction the word "enormous" points.
Run 7 put the head outside the top of the picture. Run 8 added em>a wide shot, the figure small in the frame/em> and produced two characters in one image: a cropped giant head across the top and a complete small strongman below it, as though the two instructions had been handed to two different objects. Run 9 replaced that with a sentence naming what had to be inside the picture — the whole of him, from the top of his hair to the stage under his boots, with empty backdrop above — and the framing came right. But the head arrived stretched into a tall column, because em>enormous/em> does not say which way to grow.
Run 10 named the axis: a big round ball, as wide as it is tall and as wide as his shoulders, sitting straight on the shoulders with no neck. That is the figure on the right of the first picture, and it measures 2.8 heads tall against the 4.3 to 5.0 of the numeric runs.
Even the Prompt That Worked Missed Its Own Numbers
Head 390 by 350 pixels, shoulders 610. Both comparisons in the prompt were quantities, and both came in off.
The run-10 sentence carried two measurable claims. em>As wide as it is tall/em> came back at 1.11 rather than 1.00, which is close. em>As wide as his shoulders/em> came back at 0.64 rather than 1.00, which is not. The words that did the work in that sentence were em>ball/em>, em>no neck/em>, em>stubby/em> and em>tiny/em>. The numbers rode along and were treated the same way the numbers in runs 2, 3, 5 and 6 were treated.
One thing changed that nobody asked for: when the head became a ball, the hair went away. The mustache stayed.
What Happens When It Moves
The run-10 still then went to the AI video generator as the first frame of a five-second clip, with the character told to bow and straighten up and the camera told to hold still.
Five seconds from the run-10 still. The head holds its size through the move.
The exaggeration survives the motion, which was the open question — a pushed proportion is the kind of thing a video model might quietly pull back toward normal. Measured across the frames where the head stands clear of the shoulders, its width runs between 0.99 and 1.10 of its width in the first frame, and in the final frame it is 1.01. During the deepest part of the bow the head overlaps the shoulders and this measurement cannot tell the two apart, so those frames are left out rather than reported.
What to Write Instead
Where This Stops
This is one character, one style family, and ten total runs. The six numeric-prompt trials did not produce a detectable change in this setup. Runs 7 to 10 tested silhouette wording and framing separately, so they should not be counted as numeric trials.
The shape route has its own limit, visible in runs 7 through 9: each shape word moves one thing, and anything it does not name is up for grabs — the frame, the axis, the hair. A silhouette description gets long, and that length is the price of it.
Measure Your Own Starting Point First
The single most useful run here was the dull one. Run 1 cost one credit, asked for nothing unusual, and told me what the style was going to hand me whatever I typed afterwards — a 4.4-head stocky build with wide shoulders. Every later decision was about the distance from that, and without it runs 2 and 4 would have looked like reasonable requests that the model ignored, rather than requests that were already satisfied before they were made. Generate the plain version, measure it, and write the prompt as a correction to what you got.
Test conditions (20 September 2026): the stills and clip were created with the image and image-to-video options available to the test account. Model names, credit costs, duration, resolution, and availability can change by account, region, and version; verify the current settings before recreating the workflow.
Files.base.png, one.png, two.png, loud.png, h3.png, h8.png (runs 1 to 6), shape.png, shape2.png, shape3.png, shape4.png (runs 7 to 10), all 2560×1440; bow.mp4 1280×720, 24 fps, 121 frames, 5.04 s, shown on this page as bow.gif.
The prompt blocks are reproduced as they were sent, including their spelling.
How the numbers were taken, and what was thrown away. Two instruments were built for the stills and discarded. Segmenting the whole figure from the warm backdrop counted the drop shadow and the corner vignette as part of the character. Scanning outward from the center of each row to find the body edge gave answers that disagreed with each other across the four images, because the skin tone and the backdrop are close in color. Both were replaced by reading landmarks off a coordinate grid drawn over each picture, then drawing those landmarks back on and checking them before use; the reading error is about twenty pixels. For the clip, a brightness threshold picked out only the lit half of the ball head, so the split was moved to the blue-minus-red channel, where the head sits near -50 and the backdrop and skin near -134. The head then merged with the white leotard, so instead of a bounding box the measurement reads the row-width profile of that region and takes the first local maximum from the top as the ball's widest row. That finder was pointed at known rescalings of the first frame first, returning 0.894 for a true 0.90 and 1.101 for a true 1.10, and every hit was drawn back onto its frame and checked.