AI Short Video Maker

Overview

Turn a one-paragraph brief into a 15–30 second narrative short video with original characters, Japanese-anime/cinematic style, Chinese subtitles, and synthetic voice-over. The pipeline is fully faceless, uses only local/offline tools for audio assembly, and is designed to be run inside a single WorkBuddy workspace.

Typical outputs:

• Custom birthday greeting videos for kids (e.g. "Naruto-style ninja boy, name Lele, 12 years old")
• Animated children's bedtime stories
• Short store-opening / event promotional clips
• Personalized festival / anniversary greetings

Dependencies

• ffmpeg must be available. The bundled scripts/assemble_video.py uses imageio-ffmpeg to locate the ffmpeg binary — install with pip install imageio-ffmpeg, or ensure ffmpeg is on PATH.
• Offline TTS: pip install pyttsx3 (Windows ships a zh-CN voice; macOS/Linux fall back to any system voice or a silent subtitle-only version).
• Image/video generation uses the host's ImageGen / VideoGen tools — no extra install.

Trigger Conditions

Use this skill when the user requests:

• "帮我做一个生日祝福视频"
• "用 AI 生成一段动画/短视频"
• "给我打个样：XX 风格，主角叫 XX"
• Any request to prototype a faceless AI video-delivery service on a platform like Xianyu/Taobao/Fiverr

Required Inputs

Before starting production, confirm with the user (or infer from context):

1. Recipient / protagonist (name, age, gender if relevant)
2. Theme / style (e.g. ninja/anime/fantasy/space/cute)
3. IP constraints — never generate copyrighted characters (Naruto, Disney, etc.). Use "XX-style original character" wording instead.
4. Total length target — default 15–20 seconds (3–4 clips of ~5s each + end card)
5. Language of subtitles / voice — default Simplified Chinese with Microsoft Huihui offline TTS
6. Output directory — default <workspace>/produce/<project-slug>/

Production Pipeline

Step 1: Create output directory and write storyboard

1. Create <out_dir>.
2. Write <out_dir>/分镜脚本.md with:
   • 3–4 scenes, each ~5 seconds
   • Subtitle text per scene (keep under 20 Chinese characters for readability)
   • End card text (big + small line)
   • Note: All Chinese text is added in post-production; do not ask the image/video model to render text.

Step 2: Character design image

Use ImageGen to generate a character sheet (full-body, neutral pose) matching the requested style. Requirements:

• Mention "original character" and avoid any copyrighted names/symbols.
• Use size: 1024x1536 for the character sheet.
• Save to <out_dir>/character_sheet.png.

Step 3: Scene keyframes (image-to-image locked to character sheet)

Generate one 16:9 keyframe per scene using ImageGen with image: character_sheet.png and input_fidelity: high. Save as:

• scene01_<slug>.png
• scene02_<slug>.png
• scene03_<slug>.png

If two parallel generations produce the same timestamp filename and overwrite each other, re-run the overwritten scene afterward.

Step 4: Video clips (image-to-video)

Use VideoGen in image-to-video mode for each keyframe:

• seconds: 5
• resolution: 720P
• Do NOT pass aspect_ratio — the backend rejects "16:9" and returns HTTP 400.
• Prompt should describe subtle, cinematic motion consistent with the scene.
• Save clips as clip01.mp4, clip02.mp4, clip03.mp4.

Step 5: TTS audio

Use offline TTS so the pipeline does not depend on network services:

import pyttsx3
engine = pyttsx3.init()
engine.setProperty('voice', r'HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Speech\Voices\Tokens\TTS_MS_ZH-CN_HUIHUI_11.0')
engine.setProperty('rate', 170)
engine.save_to_file(text, 's1.mp3')
engine.runAndWait()

Generate one audio file per subtitle line plus the end card line.

If the target environment lacks the Huihui voice, fall back to any zh-CN system voice or produce a silent subtitle-only version and tell the user.

Step 6: Assembly with ffmpeg

Use the bundled script scripts/assemble_video.py:

python <skill_dir>/scripts/assemble_video.py --dir <out_dir>

The script expects:

• clip01.mp4, clip02.mp4, clip03.mp4
• s1.mp3, s2.mp3, s3.mp3, s4.mp3
• scene03_<slug>.png (used for the 3-second end card background)
• msyh.ttc or simhei.ttf copied into <out_dir> (the script will try to copy it from C:/Windows/Fonts/ automatically)

The script outputs <out_dir>/<title>_成片.mp4.

Critical Pitfalls (learned from production)

Cost Benchmark

Typical 15–20 second video:

• 1 character sheet + 3 keyframes ≈ 30–50 image-gen credits
• 3 × 5-second video clips ≈ 150–300 video-gen credits
• ffmpeg assembly ≈ 0 credits

Quality Bar for Commercial Use

Before listing the service on a trading platform, verify:

• Character consistency across all scenes (no face/clothing drift)
• Subtitles render correctly in Chinese
• Audio is intelligible and synced
• Total render cost leaves healthy margin at the chosen price (recommended ¥49–199 per video)

Resources

• scripts/assemble_video.py — full ffmpeg assembly pipeline (subtitle burn-in, audio mix, end card, concat).