In this five-run test, a size adjective and detailed ratios did not reliably create a giant-scale result. The result changed only after the prompt specified the camera position and required the reference building to remain visible in the frame.
Here is the whole run, in the order it happened.
Four runs. The first three are the usual advice, followed closely. The fourth adds one sentence about the frame.
The Size Word on Its Own
The first prompt went to the AI image generator with no numbers in it, just the adjective and a consequence: a giant woman in an ordinary street, towering over the buildings. Everything else was kept deliberately plain — overcast light, wet road, parked cars, no landmark, no signage.
What came back is a woman standing in a street, photographed from road level. She is large in the frame, which is a different thing from being large. The camera is low, so she fills two thirds of the picture height; the building beside her still runs off the top edge above her head. Nothing in the shot says she is bigger than a person.
Spelling Out the Ratios, Twice
So the next prompt did what the advice says to do, and named three ratios against things in the scene.
Run 2. She is as tall as the six-storey office block beside her: the top of her head is level with the sixth-floor windows, her knee is level with the second floor, and each of her shoes is longer than the cars parked at her feet.
That came back at ordinary human size. Run 3 sent the identical prompt again, in case the first was unlucky. It also came back at ordinary human size, in a slightly different street, with a slightly different pose. Two runs of one prompt, no giant in either.
Why None of Those Three Could Be Measured
At this point the plan was to put a number on how far off each one was. That turned out to be impossible, and the reason is the finding.
Each prompt named a reference — the sixth floor, the roofline, the parked cars. In all three images, the top of that reference is outside the frame. The building is cut by the top edge above her head in each one, so there is no roofline in the picture to compare her against. The cars are the other candidate, and they sit at a different distance from the camera than she does, which makes a shoe-to-car ratio a measure of depth rather than of size.
Two instruments were built and thrown away here. Shoe length divided by the length of the nearest parked car put the three images within a hair of each other, including the one that looks visibly different. Subject height divided by the height of one window band put her at seven floors in a picture where she is plainly standing at head height to a pedestrian. Both were measuring perspective. A ratio between two objects only means something when the two objects are at roughly the same distance from the camera and both are fully inside the picture, and in these three frames neither part of that condition held.
One Sentence About the Frame
Run 4 kept the three ratios word for word and added a clause about what the photograph has to contain: the entire block from the pavement up to its roofline, the whole building inside the frame, sky above it, shot from the far side of the street.
Run 4. Same ratios as runs 2 and 3, plus one sentence describing the shot. She is now several floors tall.
She is unmistakably giant. Nothing about the size wording changed between run 3 and run 4 — the difference is the sentence about the frame, which pushed the camera back across the road and gave the picture a building to measure her against.
The Numbers Still Came in Low
Now that there is something to measure her against, the two numbered claims can be checked. The window bands on the wall she stands against were located from a brightness profile down the wall, then drawn back onto the image and checked by eye before any number was taken.
Her head sits just above the second window band. The prompt asked for the sixth floor. Her knee is inside the lowest band; the prompt asked for the second floor.
In this run, both requested comparisons landed below the requested scale: the head reached just above the second visible window band rather than the sixth, and the knee remained within the lowest band rather than the second. Treat numeric ratios as a prompt to test, not as a guaranteed output.
It is also worth saying that the framing clause was itself only partly honored. It asked for the whole building with sky above it; the roofline is visible where the building recedes down the street, and above her own head it still leaves the frame. The clause moved the camera without delivering literally what it said.
Going the Other Way: A Tiny Person and a Mug
The inverse direction is the same trend with the sizes swapped, and it is easier to check, because a mug and a person on a kitchen table are at the same distance from the camera and both fit in the shot.
Run 5. A tiny woman standing on a plain wooden kitchen table beside an ordinary white ceramic mug. The mug is more than twice her height: her head reaches only to the middle of the mug, and her shoulders are narrower than the mug's handle.
Mug 1040 px, her 870 px: a ratio of 1.20 where the prompt asked for more than 2.00. Her head clears the middle of the mug by a third of the mug's height.
This one looks convincing at a glance and is the furthest off the brief of any of the five. The same pattern as run 4 holds in the opposite direction: the relationship arrived, the quantity did not.
What Happens When It Moves
The giant frame then went to the AI video generator as the first frame of a five-second clip, with the camera told to stay locked off and one car told to drive away.
Five seconds from the run-4 still.
In this five-second test clip, the measured wall and foot-contact areas remained close to their starting size and position. This suggests that the scale held in this result, but creators should review each generated clip for camera movement, subject drift, and scale changes before exporting.
One instruction did go sideways. The prompt asked her to turn her head to look down the street, meaning along it; she spent the clip looking downward instead. Direction words that a person would read from context get read literally, so it is worth writing the one you mean — em>toward the far end of the street/em> rather than em>down the street/em>.
What to Write Instead
Where This Stops
Five images is five images. The pattern held in both directions and across two runs of one identical prompt, which is enough to act on and not enough to call a rule. The ratios were low in all five, by varying amounts, so the correction is a nudge rather than a formula — ask for more than you want and check the result, rather than multiplying by some number taken from here.
Two of the runs also differed from each other on identical input, which is worth budgeting for. A prompt that works is a prompt that works often, not a prompt that works.
The Subject of the Sentence
The five prompts here spent most of their words on the woman: how tall she is, what she is level with, how she compares to a car. The words that changed the picture were about the camera and the edges of the frame. A size illusion is a fact about a photograph, not a fact about a person, and the prompt gets further when it is written that way — describe the shot you want taken, and let the size be a consequence of it.
Test conditions (18 September 2026): this comparison used the image and image-to-video options available to the test account. Model selection, credit cost, duration, resolution, and availability may differ by account, region, and version; confirm the current settings before repeating the test.
Files.adj.png (run 1), ref.png and ref2.png (runs 2 and 3), fit.png (run 4), tiny.png (run 5), all 2560×1440; vid2.mp4 1280×720, 24 fps, 121 frames, 5.04 s,
The prompt blocks are reproduced as they were sent, including their spelling.
How the numbers were taken. Window bands were located from a brightness profile down a strip of wall clear of the subject, then drawn back onto the image and checked before being used. The two video measures use multi-scale normalized cross-correlation. Both were pointed at known transforms first: rescaling the first frame by 1.00, 0.90, 1.15 and 1.10 was recovered exactly in each case. A head patch was tried first for the video and discarded — its correlation fell from 1.00 to 0.67 once she bowed her head, so it could no longer tell a change of size from a change of pose; the foot-contact patch held at 0.98 and was used instead. Every match box was rendered back onto its frame and checked by eye before any figure here was taken from it.