netfire222/vfx

/vfx - VFX pack for talking-head reels

Big titles are drawn behind the speaker (person segmentation) and auto-placed above the head. Full Hebrew/RTL support.

Command	What it does
/vfx strike	Big word above the head -> red line strikes it -> swaps to a new word
/vfx echo	Big title pops in with outline echo copies bursting out
/vfx spotlight	Background goes dark, warm light on the speaker, flickering neon title
/vfx freeze-mark	Freeze + flash + punch-in + hand-drawn red circle around the face (+ optional label)
/vfx pinned-comment	"Reply to comment" card pops in at the top
/vfx plan	Several effects in one render
/vfx showcase	Wraps a rendered clip in a Claude-app UI: typed commands, "Edited with /vfx X" header, typing + effect sounds, ending CTA

Inputs

• The video (vertical 9:16 talking head works best) and, per effect, the exact texts and timing. Never invent texts the user did not give; ask once if missing. Timing can come from a trigger word (see words).
• Save output next to the source as <name>_vfx.mp4.

Setup (once)

• Scripts live next to this file: scripts/vfx.py, scripts/compose.py (call them by full path).
• Needs ffmpeg, Node/npm (fonts + segmentation model are fetched once with npm pack into ~/.cache/vfx) and pip install numpy pillow fonttools python-bidi ai-edge-litert "opencv-python-headless<5". Optional: pip install rembg onnxruntime (cleaner hair edges, used automatically), pip install faster-whisper (word timing).

Workflow

1. ffprobe the clip. Little headroom above the head -> titles drop to the lower third automatically; suggest reframing.
2. Timing from speech: python3 scripts/vfx.py words clip.mp4 --lang he -> {word,start,end} lines.
3. Render:

   python3 scripts/vfx.py strike in.mp4 out.mp4 --text "משעמם" --swap "מרתק" --at 2.1 --strike-at 2.8 --swap-at 3.3
   python3 scripts/vfx.py echo in.mp4 out.mp4 --text "IN-APP BROWSER" --at 5 --dur 2.5
   python3 scripts/vfx.py spotlight in.mp4 out.mp4 --text "AI SLOP" --at 13 --dur 2.5
   python3 scripts/vfx.py freeze-mark in.mp4 out.mp4 --at 9.5 --hold 1.5 --label "תסתכלו כאן"
   python3 scripts/vfx.py pinned-comment in.mp4 out.mp4 --user someone --text "..." --at 17 --dur 4
   python3 scripts/vfx.py plan in.mp4 out.mp4 plan.json      # [{"fx":"strike","text":..,"at":..}, ...]
   

   Strike variant (plan.json only): "front_first": true puts the first word ON the face, in front of the speaker, and the swap word BEHIND the head (e.g. "מלפנים" -> "מאחור"); "overlap": 0.45 hides more of it behind the head. Echo variant (plan.json): "linger": true keeps the outline copies on screen (gently breathing) until the effect ends. Options: --y (fixed title height, fraction of frame; raise it if the head hides letters), --color, freeze-mark --x --y --r to circle something else, --seg fast|quality, --crf.
4. Showcase (optional): python3 scripts/compose.py fx.mp4 final.mp4 schedule.json

   {"intro": 2.4, "cta": 3.2,
    "cmds": [{"cmd": "strike", "send": 0.0, "desc": "Say it, strike it, swap it", "sfx": [["swish", 0.9], ["pop", 1.5]]},
             {"cmd": "echo", "send": 2.75, "desc": "Big title, echo burst", "sfx": [["boom", 2.8]]}],
    "cta_lines": ["רוצים את הסקיל", "של 5 הפקודות?"], "cta_button": "תגיבו \"סקיל\""}
   

   send = clip second the command is sent (set each effect's at ~0.05s after it). First command is typed in the intro. Sounds: key, send, whoosh, swish, pop, boom, shutter, power, ding, chime.
5. Verify: grab 2-3 frames per effect (ffmpeg -ss T -i out.mp4 -frames:v 1 f.png), look at them (text readable, face visible, Hebrew right-to-left, nothing cut), fix and re-render. Then deliver with one line per effect.