Kling 4.0 video-edit template example pair, measured: frame count and aspect kept, but the input's scene cuts are not found in the output

What this is: measurements from a private, archived analysis of the example input and output videos that ship with the Comfy-Org workflow_templates Kling 4.0 video-edit template (model setting kling-4.0, resolution 720p, aspect ratio adaptive). The measured files are the example input and output at the pinned commit; their exact paths and hashes are in the citation and the source pins. The template file is cited only as the identity of that example pair: none of the measurements below are statements from it, and its own statements are quoted separately. The files were decoded privately; this record reads only the derived measurement files. The example media are not reproduced here. No video was played and no audio was listened to. What the template's own note says (author statement, not verified here): "The output inherits the input: same length, same aspect ratio (`adaptive`), same three shots." The analysis tested that sentence with a three-part rule fixed before any of this media was decoded: equal decoded frame count, scene cuts equal in number with each within 2 frames, and equal display aspect. Length: both clips have 121 decoded frames at 24 fps. Their first and last frame timestamps are 5 s apart, and each container reports 5.041667 s. The frame-count part of the rule is met. Aspect and size: both have display aspect 181/321 from a coded size of 724x1284, so the aspect part of the rule is met. The template sets 720p, yet neither clip has a 720-pixel side; the output has exactly the input's coded size. That size comparison is exploratory: it was added after the analysis ran. Shots: only cuts with an ffmpeg scene score above 0.4 are detected. The input has cuts at frames 36 and 78, which gives three shots; the output has none at that threshold, which gives one. The cut part of the rule therefore fails, and the rule as a whole is not met. A separate 64x36 grey frame-difference curve puts the output's two largest spikes at frames 78 (48.276, 8.37 times its median) and 36 (39.425, 6.83 times); the input's are at 36 (53.717, 9.33 times) and 78 (51.752, 8.99 times). Spikes may be cuts or fast motion, and transitions scoring 0.4 or less (fades, dissolves, soft cuts) are not detected, so whether the output keeps three shots is not established. Still frames: on the private contact sheets (nine evenly spaced tiles per clip), the framing scale changes in both clips between the tiles just before and just after each input cut frame. Stills show no motion and cannot locate or classify a transition, so they do not change the result above. Audio: each clip has one AAC stereo stream at 44100 Hz with a declared duration of 5.038730 s. Their loudness envelopes (first stream decoded to 8 kHz mono, 100 ms windows, 50 per clip) are not constant: the input runs from -30.656 to -9.716 dBFS and the output from -35.391 to -13.612 dBFS. The best correlation between them is r 0.9643 at lag 0, in the analysis's high envelope similarity band. An envelope band says nothing about what the sound is, or whether it was replaced, kept or carried over. Container tags: both files carry editable tags named AIGC, prompt and workflow, besides encoder and brand fields. Their value lengths are 233, 1266 and 12081 characters in the input and 233, 1150 and 3046 in the output; the values are not quoted here. Tags can be edited by anyone and prove nothing about provenance, including where the input came from. Not established: same audio identity or carry-over from the input, absolute time alignment, what the audio contains, subject or costume identity, per-shot framing similarity, preserved camera work or motion, prompt compliance and perceptual quality.

{
  "citations": [
    {
      "url": "https://github.com/Comfy-Org/workflow_templates/blob/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/api_kling4_video_edit.json",
      "note": "Source identity only: the template file at commit 0cc0fba562efe9f68e48050c2b82ed6495050bc2. The measurements come from a separate private analysis of that commit's example files input/kling4_gown_input.mp4 (4478277 bytes, sha256 8c36c3b0d2f7f20992656638c6e10798eadc03a6b29ab6b419f4b8becf348d99, git blob a7543e95754f5f211003c1e82d67b04edf08bfe3) and output/api_kling4_video_edit.mp4 (5276045 bytes, sha256 553f99b2b813d3d8813958110da8425ad7608414667d45dd7417e4ef5dab8cfa, git blob db55eb379ef40807531132ddec3f430f2bc8df47), not from statements in the template file. Originals archive 2ba303a2fcca829a6f81b0ad03a2772a262ddf59afaaf4c5ef7fea828e9d3b36; derived archive a002917b0bb4f357e455955b4e161c2294d3870cccc3293ab7de53311f2bfd11; each derived member's hash and byte range is in the source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/0cc0fba562efe9f68e48050c2b82ed6495050bc2/templates/api_kling4_video_edit.json",
      "note": "Template settings cited by byte range; the template itself is read in the published context record."
    }
  ]
}