DEV Community

Peiguan
Peiguan

Posted on

How Do You Lengthen an 8-Second Bama Bound Clip to 15s Without Inventing Motion? Freeze vs Pace Stretch (ffmpeg, Tested)

The studio will not do this for you. I build bamabound.app, and the live FAQ is explicit: the preset follows the reference timing, about eight seconds, and "You cannot type a new duration." Resolution is the only extra setting. If you need a 15-second file for a particular app, that is a post-download job.

I measured three ways of getting to 15.000 seconds without inventing new motion: hold the last frame, stretch the existing frames by changing their timestamps and duplicating them to keep 24 fps, and (as a contrast, not as a crossfade) hard-loop the clip and trim. I ran the same paths on a labeled synthetic control — a 3.000-second moving bar with a 440 Hz tone — and on the official imported field preview the homepage plays. That preview is an imported reference, not a billed generation from this studio. I did not invent customer outputs.

This is a measurement article, not a pricing article. I re-checked the pricing page today (18 November 2026, 15:31 Asia/Shanghai). A single 480p Field banner run is $7.98 (798 credits); 720p is $9.98 (998); 1080p is $16.98 (1923); 4k is $24.98 (3921); tax is added at checkout. I am quoting those only because I re-read them this afternoon.

What I started from

The official imported file (src/ref.mp4, SHA-256 prefix 9dcb91af) is 720×1280, H.264 High, yuv420p, progressive, 24 fps, 193 frames. The video stream lasts 8.041667 seconds, which is exactly 193/24. The container and the AAC LC 32 kHz stereo soundtrack both last 8.096 seconds, so there is a 0.054333-second tail after the last video frame. Colour range and colour space are unspecified. The format tag AIGC is set to {This content is generated by AI}. I re-verified those figures with ffprobe 7.1.5 today; every command I quote exited 0.

Fifteen seconds at 24 fps is 360 frames exactly. A last-frame hold therefore needs 360 − 193 = 167 cloned frames, which is 6.958333 seconds of still picture. A pace stretch needs a factor of 15 / 8.041667 ≈ 1.8652849 on the video timestamps.

The synthetic control is shorter and cleaner, so the arithmetic is obvious and the soundtrack has a known pitch. It is 72 frames at 24 fps (3.000 seconds of picture and sound), 360×640, with a red bar that walks down a green field and a 440 Hz stereo sine at 32 kHz. I built the frames with Pillow and muxed them with ffmpeg. The file is labeled synthetic.

Freeze the last frame

tpad with stop_mode=clone repeats the last decoded frame. In ffmpeg 7.1.5 the option for a frame count is stop, not a "stop_period" alias I briefly tried and watched fail. Audio is padded with silence so the two streams end together:

ffmpeg -i clip.mp4 -filter_complex \
  "[0:v]tpad=stop_mode=clone:stop=167,setpts=PTS-STARTPTS[v]; \
   [0:a]apad=whole_dur=15,atrim=0:15,asetpts=PTS-STARTPTS[a]" \
  -map "[v]" -map "[a]" -r 24 \
  -c:v libx264 -preset ultrafast -crf 18 -pix_fmt yuv420p \
  -c:a aac -ar 32000 -ac 2 out.mp4
Enter fullscreen mode Exit fullscreen mode

On the official preview, the output is 360 frames. Video, audio and container all probe at 15.000000 seconds, 24 fps. The first 193 frames compared against the source with ffmpeg's ssim filter give All = 0.987152 — that is a re-encode of moving picture, not a lossless copy. The 167-frame hold, compared against a still of source frame 192 encoded the same way, gives All = 0.996678. The late-audio window I extracted at 9.75 seconds (0.6 seconds of PCM) has RMS 0.0: the soundtrack has ended, and the pad is silence. The original music is untouched for its original 8.096 seconds. Nothing new is invented in the picture; the person simply stops.

On the synthetic control, with stop=72 and a 6.000-second target, the output is 144 frames and 6.000000 seconds on video, audio and container. SSIM of the first 72 frames against the source is 0.999989. SSIM of the hold against a still of the last source frame is 0.9995. A 0.6-second FFT window starting at 0.3 seconds still peaks at 440.0 Hz. The same window placed in the second half of the file peaks at 0.0 Hz, and its RMS is 0.0. That is the behaviour I want from a freeze: the tone stays at pitch until it ends, then there is silence, not a held note.

Pace-stretch, without interpolating new motion

setpts=FACTOR*PTS spaces the existing frames further apart. Left alone, you would have 193 frames over 15 seconds, which is about 12.87 fps. I followed it with fps=24, which duplicates frames to fill the 24 fps grid, and then trimmed to 15.000 seconds. That is a slow-motion made of holds, not a motion-interpolated 15-second take. For audio I compared two filters.

atempo=1/FACTOR changes tempo and is specified to leave pitch alone. For the official file, 1/1.8652849 ≈ 0.53612, which is inside atempo's 0.5–100 range, so it does not need to be chained.

ffmpeg -i clip.mp4 -filter_complex \
  "[0:v]setpts=1.8652848967757556*PTS,fps=24,setpts=PTS-STARTPTS,trim=0:15[v]; \
   [0:a]atempo=0.5361147753,atrim=0:15,asetpts=PTS-STARTPTS[a]" \
  -map "[v]" -map "[a]" -r 24 \
  -c:v libx264 -preset ultrafast -crf 18 -pix_fmt yuv420p \
  -c:a aac -ar 32000 -ac 2 out.mp4
Enter fullscreen mode Exit fullscreen mode

On the official preview this also probes at 360 frames and 15.000000 seconds on video, audio and container. SSIM of the first output frame against the first source frame is 0.996157. SSIM of the last output frame against the last source frame is 0.991438. I am not going to quote an FFT peak from the official soundtrack as a pitch measurement: that bed is music, not a sine, and a single peak bin is not a pitch. The pitch claim comes from the synthetic control.

On the 3.000-second, 440 Hz fixture, a factor of 2 and atempo=0.5 produce 144 frames and 6.000000 seconds of picture and container. The audio stream probes at 5.981 seconds, 0.019 seconds short of the picture — a small trim/AAC framing leftover I am leaving in the record rather than rounding away. The early FFT window peaks at 440.0 Hz, and a window in the second half also peaks at 440.0 Hz, with a non-zero RMS. The tone was slowed down and the pitch did not drop.

The other audio path is asetrate plus aresample. For a factor of 2 on a 32 kHz file that is asetrate=16000,aresample=32000. Pitch should halve. On the synthetic control, the output is a clean 6.000000 seconds on all three clocks, and both the early and the late FFT windows peak at 220.0 Hz. That is the cost of this method: the clip is longer, and it also sounds lower. I ran the same filter on the official preview (asetrate 17156, then back to 32 kHz) and it also probes at 360 frames and 15.000000 seconds, but I am not treating an FFT of that music bed as a pitch proof.

A hard loop, only as a contrast

A hard concat of the official file with itself, trimmed with -t 15, is not a freeze and is not a stretch. It also is not the crossfade loop I have written about on a different day. I include it only to show what you get if you just play the clip again.

The official output is 358 frames, video 14.958333 seconds, audio and container 15.008 seconds, and avg_frame_rate 23.9331476. It does not land on 15.000000 seconds of picture, and it puts a cut back to the opening pose in the middle of the file. The synthetic version is similar: 143 frames, video 5.958333 seconds, audio and container 6.008 seconds. If you wanted a seamless 15-second loop, you would be solving a different problem, and you would be inventing or blending frames at the join. That is outside the scope of this measurement.

How the quality numbers were taken

ffmpeg's ssim filter writes an All: value. I compared whole streams after a trim on one or both inputs, with setpts=PTS-STARTPTS so the frames line up. The official freeze "head" comparison is the first 193 frames of the freeze file against the 193-frame source. The official hold comparison is frames 193–359 of the freeze file against a 167-frame still I made from source frame 192, encoded with the same libx264 ultrafast / CRF 18 settings as the variants. I used those settings so the comparison would be between like encodes, not so I could claim a delivery encode. A slower preset would change the SSIM in the third decimal and would not change the duration figures.

The synthetic SSIMs are higher (0.999989 and 0.9995) because the fixture is a graphic with a hard-edged bar, not a photographed person in a field. I am not averaging the two tests into one "quality score." They answer different questions: did the moving part survive the re-encode, and did the hold stay like the last source frame.

PCM extracts of the official 15.000-second outputs were 15.008 seconds (480,256 samples at 32 kHz). That is 256 samples, or 8 milliseconds, past the stream duration ffprobe reports, and it is ordinary AAC framing. I quote the ffprobe stream and container durations — 15.000000 seconds on the freeze and both stretches — as the file durations, and I mention the 15.008-second PCM length only so I do not pretend the wav and the container are the same object.

I also tried holding a chosen mid-clip frame on the synthetic path, rather than the last frame, by splitting, taking one still, and padding. The duration arithmetic is the same; only the picture you freeze changes. On a real Bama Bound clip the last frame is usually the honest choice, because the motion you already paid for plays in full before the hold. Freezing an earlier smile and then jumping to it mid-action is a different edit, and I did not use it on the official file.

What this does not claim

It does not claim that a paid generation from this studio is 8.041667 seconds. That number is a measurement of the official imported preview. Paid clips follow the same preset timing in the FAQ's wording ("about eight seconds"), and I am not substituting a customer file I do not have. It does not claim that SSIM 0.987 or 0.996 is "visually lossless." It does not recommend shipping ultrafast encodes. And it does not turn the hard loop into a recipe for a seamless 15-second repeat — the official loop missed 15.000 seconds of picture by 0.041667 seconds and put a cut in the middle.

What I would actually use

If the goal is "the clip I already have, then it holds still until 15.000 seconds," freeze the last frame and pad the audio with silence. On today's official preview that produced a 360-frame, 15.000000-second file, a hold that still looks like the last source frame (SSIM 0.996678), and a silent tail. If the goal is "the same motion, played slower, for the whole 15 seconds," use setpts plus fps=24 and atempo, and expect a re-encode. On the 440 Hz control, atempo kept the tone at 440.0 Hz; asetrate dropped it to 220.0 Hz.

None of these commands change what the studio will generate. The preset still follows the reference timing, you still cannot type a new duration, and a 480p run is still $7.98 before tax unless the pricing page has moved since this afternoon. The work happens after the download, on a copy of the file, and the numbers above are only from commands that exited 0.

The scripts I used, make_lengthen_fixtures.py and lengthen_check.py, build the synthetic control, encode the variants, probe durations, run the SSIM comparisons, and take the FFT windows. I only report figures from a run that finished cleanly.

Top comments (0)