界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
开发与工程 · SKILL.md

image-series-to-video

image-series-to-video

【开发与工程】图片系列转视频工作流:将带 manifest 的 PNG 图片系列转成带讲解、字幕和验收的 MP4 视频。Workflow skill for converting a manifest-backed image series into a narrated video; use when generated PNGs or an image manifest should become an MP4, long-image explanation video, chapter video, narrated walkthrough, or social video.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: image-series-to-video description: "图片系列转视频工作流:将带 manifest 的 PNG 图片系列转成带讲解、字幕和验收的 MP4 视频。Workflow skill for converting a manifest-backed image series into a narrated video; use when generated PNGs or an image manifest should become an MP4, long-image explanation video, chapter video, narrated walkthrough, or social video."

Image Series To Video Workflow

This workflow starts only after a real image series exists. It prevents the video stage from inventing fake screenshots, CSS cards, or placeholder images.

Required Inputs

The current artifact scope must contain:

  • assets/images/*.png
  • assets/images/manifest.md

The manifest must describe one row per image with title, file, usage, scan mode, and focus guidance.

Output Contract

Final deliverables must stay inside the current task artifact scope:

  • video.config.json
  • index.html
  • assets/audio/*.mp3
  • assets/audio/bgm.wav
  • renders/*.mp4
  • ffprobe.json
  • DELIVERY.md

Phase 1: Input Gate

Validate before building video:

find assets/images -maxdepth 1 -type f -name '*.png' | sort
test -f assets/images/manifest.md

Reject the task if images are missing, empty, non-PNG, or not referenced by manifest.

Phase 2: Video Plan

Create or update video.config.json from the manifest:

  • one scene per image or chapter
  • title and caption per scene
  • scan mode and safe focus from manifest
  • voiceover text
  • inspect timestamps
  • output file name

Do not hand-write fixed scene arrays unrelated to the manifest.

Phase 3: Build

Use the repository video runner when available. Otherwise create the minimal equivalent pipeline:

  • generate index.html
  • generate voiceover audio
  • run visual acceptance snapshots
  • render MP4
  • run ffprobe

Phase 4: Acceptance

Only report completion if:

  • renders/*.mp4 exists and is non-empty.
  • ffprobe.json shows video and audio streams.
  • duration is plausible for the number of scenes.
  • sampled frames show real source images, visible captions, and no obvious overlap.

If any step fails, report the failing phase and keep partial artifacts for diagnosis.

我的收藏