Re-running can change the result, but it does not improve the source photo. A second generation may succeed by chance; changing the photo is the only way to improve framing, background simplicity, subject separation, and lighting before the request is sent.
So the same ceiling-hanging request went out six times: three from a deliberately bad photo and three from a clean one, with everything else held identical. Each result was scored against a rule written down before any of them existed.
One of the runs from the clean photo: 1280 by 720, 24 frames per second, 5.04 seconds. Shown here as frames in sequence; the full video and a GIF version are supplied alongside this document.
The two photos, and what is wrong with one of them
Both were generated for this article. The bad one breaks four of the usual rules at once, on purpose.
The cluttered photo fails on framing, on background, on company and on light: the subject is small and distant, boxes and drying laundry fill the room, a second person is walking through the doorway behind him, and the only light is a warm lamp off to one side. The clean photo is the opposite on each count — close, alone, plain wall, even daylight from the front, sharp.
How to Turn One Standing Photo Into a Ceiling-Hanging Gravity-Flip Video covers what the source photo should show and how to describe the pose. This test focuses on what happened when four of those source-photo conditions were deliberately broken.
The rule, written before the clips existed
A run counts as clean if the last frame shows all three of these conditions: the person is in contact with the ceiling, the head is lower than the feet, and one intact person appears without duplication or body distortion. A run that misses any condition is a failure. A result that meets all three but still reads awkwardly is marked marginal, and the final frames are included so readers can check the scoring.
The person in the photo is now up at the ceiling, curled and gripping it with both feet, head turned down toward the camera, one arm hanging loose, laughing with a wide open grin as if playing a prank on whoever is filming. The room around him stays exactly as it is. Locked-off tripod shot, no zoom, no push in, no dolly, no pan and no crop, and the last frame has the same framing as the first frame.
Six runs
The last frame of each run, scored against the rule above.
From the cluttered photo: two clean results and one outright failure, in which the man is simply standing in the doorway at five seconds, exactly as he was in the photograph. Nothing was attempted.
From the clean photo: two clean results and one marginal — on the ceiling and intact, but reading as lying against it rather than hanging from it.
If the marginal result is counted as usable, the cluttered photo produced two usable outputs out of three and the clean photo produced three out of three. Only four of the six results were clean. Three runs per photo are too few to establish a reliable success rate, and one non-event creates the entire difference between the two conditions. The narrower finding is that the cluttered photo produced one complete non-event, while the clean photo produced none in this test.
In this test, the failed run looked like the source photo
Change is measured against the source photo the clip was made from. Comparing a frame with itself gives 0.000, which is where the scale is anchored.
Measuring how far each final frame differs from its source photo, the one non-event separates from the other five clips: 9.8 compared with 23.2 to 40.6. In this six-clip set, the lower value marked the generation that made almost no visible change. It should not be treated as a universal failure threshold for other photos, models, or settings.
The practical check is simple: compare an unusable result with the uploaded photo. If the final frame is still very close to the source, the generation may have made little or no progress. Before rewriting the prompt, consider whether a cleaner photo would remove controllable problems in framing, background, subject separation, or lighting.
What the six runs cost
The cost lines as shown before each request, read on 15 September 2026.
The photos and clips were made through the AI image generator and AI video generator flows. These figures are for these configurations on this account and this date, and other models and durations are charged differently. The credit balance is not offered as a cross-check: it read higher at the end of the session than at the start despite the spending, so credits were added during it from something not observed here.
Price a bad photo in credits, not in annoyance
At 25 credits per run in this session, counting the marginal clip as usable gives the clean photo three usable outputs at 25 credits each. Counting only clean results, both source photos produced two clean clips, or 37.5 credits per clean result. With only three runs per condition, this does not prove a cost advantage; it shows that a non-event can consume the same credits as a usable output.
Before spending another 25 credits on a re-run, inspect both the result and the source photo. A re-run may change the outcome, but a closer subject, simpler background, single clearly separated person, and even front lighting improve the input factors you can control.
Both people were generated from text; no photograph of a real person was used. Obtain permission before using an identifiable person's photo. Keep the prompt generic rather than tying the pose to a protected film character; this test uses the terms "ceiling-hanging" and "gravity-flip" throughout.
Also not covered: repairing a weak source photo. CapCut offers an image background remover, but this experiment compared unedited source images and did not test whether background removal alone improves a gravity-flip result.
Written 15 September 2026. Both source photos and all six clips were generated in CapCut for this article; no real person appears in them. All six clips are 1280 by 720, 24 frames per second, 121 frames, 5.04 seconds. Runs were scored against the three-part rule stated above, which was fixed before the clips were generated, and all six final frames are reproduced so the scoring can be checked. Change is the mean absolute per-pixel difference between a clip's final frame and the source photo it was made from, both scaled to 1280 by 720; the zero point was verified by comparing a frame with itself, which gives 0.000. A face-detection measure was tried first and discarded: it reported three faces in the cluttered source photo, one of which was a pile of laundry. Three runs per condition is a small sample and the difference between the two conditions rests on a single failure. The findings apply to Seedance 2.0 Mini in the tested configuration; other models, settings, or later versions may behave differently.