作者分享以 5 個片段製作 35 秒寫實 AI 影片的完整藍圖
I Made a 35-Second Photorealistic AI Film With 5 Segments — Here's the Exact Blueprint
作者以 5 個獨立生成的片段製作了一部 35 秒寫實海邊日常影片，並分享讓人物、服裝與故事保持連貫的製作藍圖。流程按敍事節點分段，並在每段提示詞中原樣重複角色設定；作者建議使用最多 1–2 句短對白和細微動作，字幕與音樂則在剪輯器中添加。完成前逐段檢查人物、服裝、道具狀態及生成文字，失敗時只重做該段；作者指出，接續前一段尾幀在其測試中較容易造成臉部和服裝漂移。
Most AI video tools stay stable for about 5–10 seconds per generation. So how do you make a 35-second film that actually holds together — same person, same outfit, one continuous story?
I shipped one: a photorealistic seaside day-in-the-life film, 35 seconds, built from 5 independently generated segments. Here's the exact blueprint I used. It works because the segments connect through discipline, not through luck.
The 5-segment blueprint
Cut on narrative beats, not on time. My film's five beats:
- Leave home — the person steps out in the morning light
- Arrive — reaching the seaside town
- Buy / event — a small shop visit, the day's one real event
- Return — heading back as the light turns
- Quiet ending — a still, held moment to close
Each segment is generated independently, then stitched in an editor. Merging 1–2 shots per segment is fine; chaining from the previous segment's last frame is not — in my tests, tail-frame chaining drifts faces and outfits the worst.
The three rules that hold it together
1. One character lock block, pasted verbatim everywhere.
Write the person once — hair, eyes, outfit, accessories, in a fixed order — and paste that exact block into every segment prompt. The block stays frozen; only the action and scene change per segment. Miss one item and that item is the one that drifts.
2. Dialogue: 1–2 shortest lines, max.
Long dialogue in generated video turns to mush. One or two short lines carry the scene; everything else is visual. And never ask the model for subtitles — add text and music in your editor, not in the model.
3. Micro-movement, not big action.
For photorealism, small wins: a slow blink, hair stirred by a faint breeze, a hand reaching for a door. The smaller the motion, the better the look survives generation.
The negative-constraint tail (attach to every segment)
no subtitles, no text, no watermark, no logos, normal hands (five fingers),
no distorted face, no face change
One segment, assembled
[person lock block], walking down a narrow seaside street in the morning,
carrying a paper bag from the bakery, light steps, side tracking shot,
gulls in the distance, natural light, photoreal, no subtitles, no text,
no distorted face, no face change
The pre-delivery frame check
Before I call a film done, every segment gets the same five questions: same face? Hair and accessories complete? Outfit colors unswapped? Prop states continuous (a bottle cap may only be open after the segment where it opens)? Any stray generated text? If a segment fails, I re-run that segment alone — never the whole film.
Where this comes from
This is one of the 36 templates in my AI Short-Video Prompt Pack Vol.1 — the full system behind 30+ shipped AI shorts: segmented generation for 15–35s films, the character lock block, image-to-video style replication, and the frame-check routine, each template annotated with the tool, aspect ratio and segmentation it was verified with. Bilingual EN / 中文 edition.
- 📦 Full pack on itch.io ($4.99): https://zzstudio.itch.io/ai-video-prompt-pack-vol1
- 🆓 Free starter (9 templates, same system): https://zzstudio.itch.io/ai-video-prompt-starter-pack
- 🆓 Methodology chapter free on GitHub: https://github.com/zz583/ai-video-prompt-pack
If you try the 5-segment blueprint, tell me in the comments how your film held together.
來源：DEV Community · dev.to