video-still-animator

name	video-still-animator
description	Turn a single still image (PNG/JPG) into a short MP4 with a slow Ken-Burns zoom and a silent audio track. Pure ffmpeg wrapper. Designed as the on_failure substitute for AI video-gen steps that get blocked by content moderation: when seedance refuses, this skill emits a valid replacement clip from the already-generated still so a downstream merge can still produce a complete deliverable.
provenance	{"origin":"agentos-original","license":"MIT"}
metadata	{"agentos":{"risk":"medium","capabilities":["filesystem-write","process-control"],"requires":{"bins":["ffmpeg"],"anyBins":["python","python3"]}}}
entrypoint	{"command":"python {baseDir}/scripts/animate.py","args":["--input","{{ with.input_image }}","--output","{{ with.output_path }}","--duration","{{ with.duration | default(5) }}","--width","{{ with.width | default(720) }}","--height","{{ with.height | default(1280) }}","--fps","{{ with.fps | default(24) }}","--zoom-rate","{{ with.zoom_rate | default(0.0015) }}"],"parse":"text","timeout":120}

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:)

key	required	default	notes
input_image	yes	—	PNG or JPG path.
output_path	yes	—	Output .mp4 path. Parent dir created if missing.
duration	no	5	Clip length in seconds.
width	no	720	Output width. Match the merge step's pipeline.
height	no	1280	Output height. 720x1280 = 9:16.
fps	no	24	Output frame rate.
zoom_rate	no	0.0015	Per-frame zoom delta; 0.0015 over 5s ≈ 1.18× final zoom.

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:

- 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.