Video digest · split clips by timeline
Updated at:
Upload a video, Qwen3.8-flash vision splits it by timeline into segments with per-segment summaries, timestamps, overall summary, and clip count.
Hand the video to the model; it splits by timeline into segments and summarizes each, with an overall summary and clip count.
Note
Disclaimer: All content on this page is AI-generated, for reference only, and does not constitute any performance guarantee. Actual results and Credits consumption are subject to the console "Usage Details".This walkthrough uses a 36-second montage: meeting recordings and compilations don't need to be watched from start to finish—let qwen3.8-flash split by timeline and describe each segment. The input is a 7-segment, ~5-second-each montage stitched by ffmpeg with a known order—the ground truth—so timestamp accuracy can be verified directly.
Call chain: qwen3.8-flash (video base64 input → timestamp segments + per-segment summary + overall summary). Vision understanding, video input.
Example input
Input material:
Steps: upload video and request timeline output
Send the video along with the prompt below to qwen3.8-flash:
<video_url: base64 464 KB>
This is a video compilation. List each segment by timeline in the format: mm:ss — visual content and key information. Append an overall summary and a segment count at the end.
View measured output · 7-segment timeline + overall summary (full)
**Segment timeline**
- **00:00–00:05** — 3D cartoon: on a sunny meadow, a white rabbit with a pink bow, a meerkat holding a sunflower, a squirrel in a yellow scarf, and a hedgehog dance together; the meerkat raises the sunflower high. Key info: wholesome cute-animal animation (watermark "Happy Home").
- **00:05–00:10** — Product still life: a frosted glass serum bottle with a gold dropper cap, blank label, on a marble counter with green leaves, warm beige background. Key info: AI-generated skincare ad visual, natural minimalist aesthetic.
- **00:10–00:15** — Night portrait: a young woman in a beige sweater stands by a window, gazing at the city nightscape with phone in hand, then puts the phone away and crosses arms to continue gazing; a warm lamp glows on the sill. Key info: solitude and waiting in a city night.
- **00:15–00:20** — Family warmth: a silver-haired grandmother with glasses sits on a sofa in a warm-lit living room, smiling and talking into a phone, seemingly on a family video call. Key info: elderly family video-call scene.
- **00:20–00:25** — Handmade food close-up: weathered hands wrap dumplings on a floured wooden board, neat dumplings, a bowl of filling, and a bamboo steamer nearby. Key info: traditional handmade dumpling wrapping.
- **00:25–00:30** — Birthday table: a steaming bowl of longevity noodles with a fried egg, a small cake with two lit candles, blue-and-white porcelain plate and chopsticks, evening kitchen background. Key info: cozy birthday/longevity meal.
- **00:30–00:35** — Cinematic period: a young man in a dark trench coat stands on an ancient city wall hung with lanterns, then turns and walks away along a stone path, cold tones. Key info: solitary, literary mood of walking on an old city wall.
**Overall summary**
This is a ~35-second, 7-segment AI-generated short video compilation, each ~5 seconds, spanning cute animation, skincare ad, urban night solitude, elderly family call, handmade dumplings, longevity-noodle birthday, and cinematic old-wall walk. The style shifts from adorable to warm slice-of-life, closing on a cold literary note—a "cute + life + emotion" montage collection (all segments except the first carry an "AI-generated" watermark).
**Segment count**: **7** segments total.
Notes
- Consumption: This playbook uses vision understanding (qwen3.8-flash, video input). Tokens converted from video count toward input; longer videos mean higher input-side cost, which dominates for long videos.
- Measured result: All 7 segments identified correctly; order and per-segment content matched. Detail recovery down to "pink bow" and "gold dropper cap" level, and it recognized the "AI-generated" watermark on screen.
- Note · timestamp precision: Model timestamps are on 5-second marks (00:00–00:05 through 00:30–00:35), while each segment is slightly over 5 seconds and total is 36 seconds. Deviation accumulates, ~1 second off by the last segment. Good enough for "which segment says what," not for precise edit points.
- Note: The test material is a 36-second short video; long videos of tens of minutes are untested. Don't directly apply this playbook's expectations to long material; cut long footage into segments and summarize separately.
- Material statement: The montage in this example is AI-generated and stitched; frames carry an "AI-generated" watermark.
Is this page helpful?