hypersocialinc/agent-skills · Archived

transparent-video

Generate a short looping transparent video (HEVC `hvc1` with alpha) from a single still image. Pipeline: SeedDance 2 image-to-video with optional end-frame anchoring for a seamless loop → ffmpeg split → BiRefNet v2 Matting per frame in parallel → ffmpeg encode HEVC with alpha. Output is a small (~1-2 MB) `.mov` that plays natively in iOS `AVPlayer`, web video tags, and any HEVC-alpha-capable runtime. Use when the user wants to animate a static character/hero image into an empty-state loop, ambi…

First seen May 23, 2026

Installation

$ npx skills add hypersocialinc/agent-skills --skill transparent-video

Summary

  • Generate a short looping transparent video (HEVC `hvc1` with alpha) from a single still image.
  • Pipeline: SeedDance 2 image-to-video with optional end-frame anchoring for a seamless loop → ffmpeg split → BiRefNet v2 Matting per frame in parallel → ffmpeg encode HEVC with alpha.
  • Output is a small (~1-2 MB) `.mov` that plays natively in iOS `AVPlayer`, web video tags, and any HEVC-alpha-capable runtime.
  • Use when the user wants to animate a static character/hero image into an empty-state loop, ambient onboarding hero, badge animation, or any short transparent video.
  • Don't use for chroma-key/green-screen footage (BiRefNet is ML segmentation, not color-keying), long videos (cost scales linearly with frame count), or audio.

Stronger alternatives

This repository is archived — consider an actively maintained alternative.

Similar popular skills

Related neighbors and high-traction skills in the same topics — useful to compare before installing.

Also in this package

Other skills from hypersocialinc/agent-skills · top by installs.

npx skills add hypersocialinc/agent-skills

Browse all from hypersocialinc/agent-skills

More details

Agent compatibility

Declared targets from SKILL.md / docs. Unmarked agents are not listed — the skill may still install via the CLI.

Claude Code Not declared
Cursor Not declared
Codex Not declared
GitHub Copilot Not declared
Windsurf Not declared
Gemini CLI Not declared
Cline Not declared
OpenCode Not declared

Repository health

Stars 1
Default branch main
Open issues 0
Status Archived

Package contents

Files included with this skill beyond the listing page.

  • skill md SKILL.md 4,069 B
  • docs SUMMARY.md 757 B

History

  1. First seen on skills.sh
  2. First recorded snapshot · 6 installs

SKILL.md

Transparent Video Loop

Workflow

  1. Confirm prerequisites. The user needs genmedia (fal.ai CLI, configured with API key), ffmpeg (Homebrew ffmpeg with hevc_videotoolbox — bundled on macOS), and a still image they want to animate. Check with genmedia setup --status and which ffmpeg.
  1. Pick the inputs. Ask the user (or infer from context) for:

- still image — local path or URL. Subject on transparent or solid background, ideally already isolated. If it's a complex scene with subject + environment, the loop will animate the whole frame. - motion prompt — what subtle ambient motion to apply. Keep it small: "characters breathe gently, hand waves slowly, leaves rustle." Big camera motion or scene changes break the closed-loop seam. - closed loop or open? — recommend closed for ambient UI loops; open is fine if the clip plays once or you don't mind a tiny seam. - output path — where to save the final .mov.

  1. Run the orchestrator. scripts/transparent-loop.sh handles the full pipeline. Recommended invocation for a closed UI loop:

``bash scripts/transparent-loop.sh \ --image ./hero.png \ --prompt "Subtle ambient idle motion that returns to the starting pose. Static camera." \ --output ./hero_loop.mov \ --duration 5 \ --closed-loop ``

The script uploads the still to fal if it's a local path, runs SeedDance 2 fast image-to-video with endimageurl = image_url (closed-loop mode), splits frames with ffmpeg, processes each frame through BiRefNet v2 Matting in parallel (10 workers by default), then encodes HEVC alpha at 720×720 24fps.

  1. Inspect the output before shipping. Open the .mov in QuickTime to verify: the loop is seamless, the alpha is clean (no halo, no missing characters), the file size is reasonable (1-3 MB for a 5s 720² clip). If a character disappeared, see the BiRefNet failure mode in references/troubleshooting.md — usually means the still had thin/translucent details the matting model can't recover.
  1. Wire into the target app. iOS uses AVPlayer directly; the .mov plays with alpha out of the box if playerLayer.pixelBufferAttributes is set to kCVPixelFormatType_32BGRA. Web uses <video> with the same .mov. See references/playback.md for the per-platform integration snippet.

Reference

Read these in order when implementing or debugging:

  • references/pipeline.md — every command, every flag, what each step does, cost breakdown.
  • references/troubleshooting.md — failure modes I hit and how to recognize them (hypervideo chroma-key trap, ProRes oversize trap, seam without endimageurl, halo without refine_foreground).
  • references/playback.md — minimal iOS / web / Android playback integration with alpha preserved.

Scope

This skill is for short transparent animation loops, typically 3–10 seconds. For longer videos, the per-frame BiRefNet cost scales linearly (~$0.005/frame, ~$0.60 for a 2-min clip at 24fps) and the resulting .mov grows quickly. For real-time effects, transparent live video, or scenes with complex semi-transparency (glass, smoke, hair on cluttered backgrounds), a dedicated VFX matting tool is better than this open-source pipeline.