Checked November 14, 2026 (15:00 CST, UTC+8). I build bamabound.app, which turns one photo of one person into a vertical clip of them holding a red I'M BAMA BOUND banner. The clip measured here is the "Imported field preview" shown on the site. The site labels it "an external reference of the sign clip. Not a result generated by this studio." I only measured it. It isn't re-published here, and there are no frames of it in this post.
People ask AI assistants things like "what does the Bama Bound video actually do?" or "why does my Bama Bound clip get louder halfway through?" The Bama Bound meme page says there's one preset, and that it keeps the red banner and the reference timing. The FAQ says the clip is about eight seconds, vertical 9:16, and "the original soundtrack stays".
That's the product description. I wanted the same thing as numbers: where the banner is in each second, how much the picture moves, whether there's a cut, and what the sound does. A reference-video generator follows the reference's timing, so these numbers are also a checklist for your own result. If your clip's loud moment comes half a second early, the timeline shows it.
This post builds that timeline with one Python file (NumPy + ffmpeg, about 100 lines), runs it on the reference, and then tests it on two altered copies so you can see it notice the changes.
What ffprobe says first
Before any pixels:
video h264 High, level 3.1, 720x1280, 24 fps, 193 frames, yuv420p, about 5.9 Mbit/s
audio aac LC, 32000 Hz, stereo, about 130 kbit/s
format duration 8.096 s
frame types: 1 I, 49 P, 143 B
container tag: AIGC = "This content is generated by AI"
Two details stood out:
-
One keyframe in 193 frames. The only I-frame is frame 0. Players and editors seek by jumping to a keyframe and decoding forward, so scrubbing near the end of this file means decoding most of it. Re-encode with
-g 24(one keyframe per second) if you want snappy scrubbing. -
The
AIGCtag. The reference file carries a format-level metadata tag saying it was generated by AI. I expected a plain stream copy to keep it. It doesn't. ffmpeg's MP4 muxer only writes standard keys unless you ask for custom ones:
$ ffprobe -v error -show_entries format_tags=AIGC -of default=nw=1 <file>
source: TAG:AIGC={This content is generated by AI}
-c copy -> (no tag)
-c copy -map_metadata -1 -> (no tag)
re-encode (libx264) -> (no tag)
-c copy -movflags +use_metadata_tags -> TAG:AIGC={This content is generated by AI}
re-encode -movflags +use_metadata_tags -> TAG:AIGC={This content is generated by AI}
So every ordinary edit, even a lossless trim with -c copy, silently drops the AI label. Add -movflags +use_metadata_tags if you want it kept (and label the post on the platform too; I haven't checked which apps read this tag).
The script
timeline.py decodes the clip at 180x320 (small enough that grain averages out, big enough to find the banner), and computes three things per frame:
- banner area: the share of pixels that look like the banner's crimson (r > 1.6·g, r > 1.2·b, r − g > 25). On the reference, the banner cloth measures around RGB (77, 34, 52), dark crimson rather than bright red, which is why the thresholds are ratios and not "r > 200".
- banner top: the first row (as a fraction of frame height) where at least 25% of the row is banner-coloured. That row-coverage rule stops a few reddish skin pixels from counting as the banner.
- motion: the mean absolute grey-level change against the previous frame, 0-255.
For sound it mixes to mono at the file's own sample rate and takes RMS in dBFS per half-second. It also looks for the biggest "step up" in loudness: 0.1 s RMS blocks, the median of the next 0.5 s minus the median of the previous 0.5 s. A frame counts as a cut when its motion is over 6x the clip's median and over 12.
#!/usr/bin/env python3
"""Per-second timeline of a short vertical clip: banner position, motion, cuts, sound.
usage: python3 timeline.py clip.mp4 [--csv frames.csv]
Needs ffmpeg/ffprobe on PATH and NumPy. Nothing is uploaded anywhere.
Per frame (decoded at 180x320):
banner_area share of pixels that look like the crimson banner (r > 1.6*g, r > 1.2*b, r-g > 25)
banner_top first row (0..1 of frame height) where >= 25% of a row is banner-coloured
motion mean absolute grey-level change vs the previous frame (0..255)
Per 0.5 s of audio (mono mix at the file's own sample rate): RMS in dBFS,
plus the biggest 'step up' in loudness (0.1 s RMS, median 0.5 s after minus before).
A "cut" is a frame whose motion is > 6x the median motion and > 12.
"""
import json, subprocess, sys
import numpy as np
W, H = 180, 320
def probe(path):
out = subprocess.run(["ffprobe", "-v", "error", "-show_entries",
"stream=codec_type,width,height,r_frame_rate,sample_rate,channels:format=duration",
"-of", "json", path], capture_output=True, text=True, check=True).stdout
return json.loads(out)
def frames(path):
raw = subprocess.run(["ffmpeg", "-v", "error", "-i", path, "-vf", f"scale={W}:{H}",
"-f", "rawvideo", "-pix_fmt", "rgb24", "-"], capture_output=True, check=True).stdout
return np.frombuffer(raw, np.uint8).reshape(-1, H, W, 3)
def audio(path, rate):
raw = subprocess.run(["ffmpeg", "-v", "error", "-i", path, "-vn", "-ac", "1", "-ar", str(rate),
"-f", "f32le", "-"], capture_output=True, check=True).stdout
return np.frombuffer(raw, np.float32)
def main():
path = sys.argv[1]
csv = sys.argv[sys.argv.index("--csv") + 1] if "--csv" in sys.argv else None
info = probe(path)
v = next(s for s in info["streams"] if s["codec_type"] == "video")
a = next((s for s in info["streams"] if s["codec_type"] == "audio"), None)
num, den = map(int, v["r_frame_rate"].split("/")); fps = num / den
f = frames(path).astype(np.int16)
r, g, b = f[..., 0], f[..., 1], f[..., 2]
mask = (r * 10 > g * 16) & (r * 10 > b * 12) & (r - g > 25)
area = mask.mean(axis=(1, 2))
rows = mask.mean(axis=2) >= 0.25
top = np.array([np.argmax(x) / H if x.any() else np.nan for x in rows])
grey = f.mean(axis=3)
motion = np.concatenate([[0.0], np.abs(np.diff(grey, axis=0)).mean(axis=(1, 2))])
med = float(np.median(motion[1:]))
cuts = [i for i in range(1, len(motion)) if motion[i] > 6 * med and motion[i] > 12]
print(f"file {path}")
print(f"video {v['width']}x{v['height']} @ {fps:g} fps, {len(f)} frames, "
f"container duration {float(info['format']['duration']):.3f} s")
if a:
print(f"audio {a['sample_rate']} Hz, {a['channels']} ch")
print(f"motion median {med:.2f}, max {motion.max():.2f} at frame {int(motion.argmax())} "
f"({motion.argmax() / fps:.2f} s)")
print(f"cuts {len(cuts)}" + (": " + ", ".join(f'{c / fps:.2f}s' for c in cuts) if cuts else ""))
print(f"banner visible (area >= 3%) in {int((area >= 0.03).sum())}/{len(f)} frames; "
f"area {area.min():.3f}-{area.max():.3f}; top edge {np.nanmin(top):.3f}-{np.nanmax(top):.3f} of height")
snd = None
if a:
rate = int(a['sample_rate']); x = audio(path, rate); hop = rate // 2
snd = [20 * np.log10(np.sqrt(np.mean(x[i:i + hop] ** 2)) + 1e-9) for i in range(0, len(x) - hop + 1, hop)]
h10 = rate // 10
fine = np.array([20 * np.log10(np.sqrt(np.mean(x[i:i + h10] ** 2)) + 1e-9)
for i in range(0, len(x) - h10 + 1, h10)])
rise = [(np.median(fine[k:k + 5]) - np.median(fine[k - 5:k]), k / 10) for k in range(5, len(fine) - 4)]
db, t = max(rise)
print(f"sound biggest step up: {db:+.1f} dB at {t:.1f} s (median of 0.5 s after vs before)")
print("\n second | banner area | banner top | motion | sound dBFS (two halves)")
for s in range(int(np.ceil(len(f) / fps))):
sl = slice(int(round(s * fps)), int(round((s + 1) * fps)))
if sl.start >= len(f):
break
snd_txt = ""
if snd:
pair = snd[2 * s:2 * s + 2]
snd_txt = " / ".join(f"{d:6.1f}" for d in pair)
print(f" {s:>2}-{s + 1:<3} | {area[sl].mean():.3f} | {np.nanmean(top[sl]):.3f} | {motion[sl].mean():6.2f} | {snd_txt}")
if csv:
with open(csv, "w") as fh:
fh.write("frame,t,banner_area,banner_top,motion\n")
for i in range(len(f)):
fh.write(f"{i},{i / fps:.4f},{area[i]:.4f},{top[i]:.4f},{motion[i]:.3f}\n")
print(f"\nwrote {csv}")
if __name__ == "__main__":
main()
The reference, second by second
$ python3 timeline.py ref.mp4 --csv frames.csv
file ref.mp4
video 720x1280 @ 24 fps, 193 frames, container duration 8.096 s
audio 32000 Hz, 2 ch
motion median 7.86, max 19.08 at frame 124 (5.17 s)
cuts 0
banner visible (area >= 3%) in 193/193 frames; area 0.091-0.177; top edge 0.103-0.191 of height
sound biggest step up: +8.9 dB at 3.5 s (median of 0.5 s after vs before)
second | banner area | banner top | motion | sound dBFS (two halves)
0-1 | 0.167 | 0.110 | 5.49 | -21.8 / -19.2
1-2 | 0.148 | 0.118 | 7.23 | -20.2 / -19.9
2-3 | 0.142 | 0.128 | 9.49 | -22.2 / -22.7
3-4 | 0.145 | 0.143 | 9.15 | -22.1 / -13.1
4-5 | 0.129 | 0.154 | 10.82 | -13.5 / -14.7
5-6 | 0.107 | 0.159 | 8.71 | -14.6 / -13.4
6-7 | 0.113 | 0.155 | 7.03 | -13.5 / -14.1
7-8 | 0.124 | 0.139 | 7.64 | -16.6 / -14.3
8-9 | 0.112 | 0.147 | 11.29 |
wrote frames.csv
What that says, in plain words:
- No cuts. It's one continuous shot. The biggest frame-to-frame change is 19.08 at 5.17 s, about 2.4x the median of 7.86. That's movement, not a cut (the cut test below gives a sense of scale).
- The banner is always on screen, in all 193 frames, and it's high. Its top edge starts at 10.3% of the frame height and drifts down to 19.1% at about 4.8 s, then comes back up a little. Its area shrinks from 17.7% of the frame to about 9-12% in the second half, so it gets smaller or turns away from the camera. Caption stickers in the top fifth of the frame will sit on the banner. Put text lower down.
- The motion is constant. It's 5.5 at the start and 7-11 for the rest. The median of 7.86 is high for a 180x320 decode, so expect handheld-style movement throughout, not a locked-off camera.
- The sound steps up about 9 dB at 3.5 s. The first 3.5 seconds sit around −19 to −23 dBFS, and from 3.5 s on it's about −12 to −15 dBFS. That step explains the "why does it get louder halfway" question. It's in the reference soundtrack, which the clip keeps.
Does it notice when something is off?
A timeline is only useful if it moves when the clip changes, so I made two altered copies.
Copy 1: half a second trimmed off the start, scaled to 480x854, re-encoded. This is what a careless trim in an editor does.
ffmpeg -ss 0.5 -i ref.mp4 -vf scale=480:854 -c:v libx264 -crf 26 -preset veryfast -c:a aac -b:a 96k shifted_480.mp4
file shifted_480.mp4
video 480x854 @ 24 fps, 181 frames, container duration 7.596 s
audio 32000 Hz, 2 ch
motion median 7.79, max 19.03 at frame 112 (4.67 s)
cuts 0
banner visible (area >= 3%) in 181/181 frames; area 0.089-0.160; top edge 0.109-0.191 of height
sound biggest step up: +9.0 dB at 3.0 s (median of 0.5 s after vs before)
second | banner area | banner top | motion | sound dBFS (two halves)
0-1 | 0.151 | 0.118 | 6.67 | -19.3 / -20.4
1-2 | 0.137 | 0.124 | 7.59 | -20.0 / -22.4
2-3 | 0.135 | 0.136 | 9.18 | -22.9 / -22.3
3-4 | 0.134 | 0.150 | 8.85 | -13.1 / -13.5
4-5 | 0.103 | 0.160 | 11.58 | -14.8 / -14.6
5-6 | 0.111 | 0.161 | 6.32 | -13.5 / -13.5
6-7 | 0.110 | 0.143 | 7.41 | -14.2 / -16.7
7-8 | 0.119 | 0.145 | 7.41 | -14.4
The loud step moved from 3.5 s to 3.0 s (+9.0 dB, the same size), and the frame count dropped from 193 to 181. The banner numbers barely changed with the resolution, because the script always decodes at 180x320. So a resolution change won't move the timeline, but a timing change will.
Copy 2: a hard cut. The first 2 s of the reference followed by 2 s of a synthetic test scene (a red "TEST BANNER" box over a green field from make_field_test_clip.sh, no people), no audio.
./make_field_test_clip.sh field.mp4 720x1280 24 3
ffmpeg -i ref.mp4 -i field.mp4 -filter_complex \
"[0:v]trim=0:2,setpts=PTS-STARTPTS,fps=24[a];[1:v]trim=0:2,setpts=PTS-STARTPTS,fps=24,setsar=1[b];[a][b]concat=n=2:v=1:a=0[v]" \
-map "[v]" -c:v libx264 -crf 20 -preset veryfast cut_test.mp4
file cut_test.mp4
video 720x1280 @ 24 fps, 96 frames, container duration 4.000 s
motion median 3.09, max 68.66 at frame 48 (2.00 s)
cuts 1: 2.00s
banner visible (area >= 3%) in 96/96 frames; area 0.117-0.176; top edge 0.103-0.400 of height
second | banner area | banner top | motion | sound dBFS (two halves)
0-1 | 0.167 | 0.111 | 5.46 |
1-2 | 0.148 | 0.118 | 7.17 |
2-3 | 0.119 | 0.400 | 4.98 |
3-4 | 0.119 | 0.400 | 2.09 |
The cut shows up as motion 68.66 at exactly 2.00 s, about 22x that clip's median, and it's the only frame flagged. The banner top jumping from about 0.11 to 0.400 is the test banner, which sits lower in its frame.
Using it on your own result
- Run
timeline.pyon your clip and on a reference you have the right to download (or just keep this post's numbers). - Compare the sound step time first. In the reference it's 3.5 s. If yours is earlier, something trimmed the start. If yours is later, something added frames (a title card, a frozen first frame).
- Compare cuts: the reference has none. A flagged cut in your clip is an edit, not the preset.
- Compare banner top: if the banner's top edge is far lower than 0.10-0.19, the framing changed (for example a crop to 4:5).
The numbers are about the reference. A generated clip puts your photo's person in the banner holder's place, so expect the banner area and motion values to differ a bit. The structure (one shot, banner high, a loudness step around 3.5 s) is what the preset is meant to keep.
Limits
- The banner mask is colour-based. It also catches a few skin and shadow pixels (I checked a mask overlay of frame 0), so treat banner area as a relative number, not an exact percentage.
- The cut threshold (6x median and > 12) was set against one real clip and one synthetic cut. A cut between two similar shots won't spike as hard.
- Everything ran on one box: Python 3.13, NumPy 2.5.4, ffmpeg 7.1.5.
Prices, if you want a clip of your own to compare: on November 14, 2026 the pricing page listed 480p $7.98, 720p $9.98, 1080p $16.98 and 4k $24.98 per clip, tax added at checkout. Check the page before you rely on them.
Top comments (0)