Video Still Animator

video-still-animator — Ken-Burns fallback clip from a still

Produces a short MP4 from a single PNG/JPG using ffmpeg's zoompan filter, so a downstream merge step still has a clip for every shot when an upstream AI video generation step gets blocked.

Why this skill exists

seedance-2-prompt (and any platform-moderated video model) can refuse individual requests for reasons that have nothing to do with the prompt's narrative quality:

• input photo contains a recognisable real person face
• output audio gets flagged by the safety classifier
• the moderation model's similarity threshold drifts day to day

In meta-short-drama we always have a fresh PNG on disk before the seedance call runs, so the cheapest "the show must go on" fallback is to turn that PNG into a Ken-Burns clip and feed it to the merger. The clip won't have the camera motion the prompt asked for, but the character, framing, and timing will still match the script.

Inputs (with:)

Output specs

• H.264 video at the requested resolution + fps
• AAC silent audio track (so merge steps that demux/remux audio do not trip on a missing stream)
• +faststart MP4 (web playback friendly)

Dependencies

• ffmpeg ≥ 5.0 on PATH (or pass --ffmpeg-path directly).

On Windows the script also probes the winget Gyan.FFmpeg install path, Scoop, Chocolatey, and C:\Program Files\ffmpeg\bin before giving up.

Use as on_failure substitute

In a meta-skill DAG, hang it off the real video step:

yamlCopy code
- id: shot1_video
  kind: skill_exec
  skill: seedance-2-prompt
  on_failure: shot1_video_fallback
  with: { ... }

- id: shot1_video_fallback
  kind: skill_exec
  skill: video-still-animator
  with:
    input_image: "{{ inputs.workspace_dir }}/.../1_shot.png"
    output_path: "{{ inputs.workspace_dir }}/.../1_shot.mp4"

The fallback must be a standalone step (no depends_on, no own on_failure) per the meta-skill engine rules. The PNG must already exist on disk by the time the parent step fails — that's true here because the image step always runs before the video step.

Limits

• Static framing only. The zoom is uniform centre-anchored; no pan, no rotation, no parallax. For richer fallback motion, generate a fresh set of stills and call this multiple times.
• No real audio. The intent is a placeholder that survives the merge; add a real soundtrack downstream if needed.
• ffmpeg version drift: very old ffmpeg (<4) may not accept the exact zoompan filter syntax; install ≥ 5.