| name | workflow-vod-post-production |
| description | Traditional VOD finishing — transcode to mezzanine + proxies, cut/trim/concat, color grade (LUT / OCIO / manual), stabilize handheld, denoise, retime (slow-mo / reverse / speed ramps), chromakey, lower thirds, burn subtitles, loudness normalize, thumbnails/sprite sheets, chapter metadata, and final delivery H.264/AV1 for YouTube/social/archive. Use when the user says "finish my YouTube video", "post-production", "color grade and stabilize", "convert to vertical TikTok", "batch wedding videos", "sprite sheet for scrubbing", or standard VOD finishing. |
| argument-hint | ["source"] |
Workflow — VOD Post-Production
What: Finish a video between editorial and distribution. Transcode, color, retime, caption, audio, package. Not live. Not broadcast master. Not AI-restoration. Just classic VOD finishing.
Skills used
ffmpeg-transcode, ffmpeg-hwaccel, ffmpeg-cut-concat, ffmpeg-speed-time, ffmpeg-video-filter, ffmpeg-stabilize, ffmpeg-denoise-restore, ffmpeg-lut-grade, ffmpeg-ocio-colorpro, ffmpeg-hdr-color, ffmpeg-chromakey, ffmpeg-compose-mask, ffmpeg-lens-perspective, ffmpeg-subtitles, ffmpeg-audio-filter, ffmpeg-audio-fx, ffmpeg-audio-spatial, media-ffmpeg-normalize, ffmpeg-frames-images, ffmpeg-metadata, media-moviepy, media-handbrake, media-batch, media-cloud-upload, ffmpeg-probe, media-mediainfo, ffmpeg-quality.
Pipeline
Step 1 — Transcode to mezzanine
ProRes 422 HQ (Apple) or DNxHR HQ (Avid / cross-platform). Never cut from camera original.
Step 2 — Proxies for offline edit
Hardware-accelerated H.264 (NVENC / QSV / VideoToolbox). Real-time or faster on modern chips.
Step 3 — Cut / trim / concat
ffmpeg-cut-concat:
trim --mode stream-copy — FAST, keyframe-snapped.trim --mode reencode — frame-accurate.concat --mode copy — same codec.concat --mode filter — different codecs.
Step 4 — Color grade
ffmpeg-lut-grade for .cube/.3dl/Hald CLUT. ffmpeg-ocio-colorpro for OCIO/ACES. Manual eq, curves, colorbalance filters via ffmpeg-video-filter.
Step 5 — Stabilize
ffmpeg-stabilize two-pass (vidstabdetect → vidstabtransform) for best quality; deshake single-pass for speed.
Step 6 — Denoise
ffmpeg-denoise-restore: nlmeans for film grain, bm3d heavier preserves detail, hqdn3d general fast.
Step 7 — Scale / crop / reframe
scale, cropdetect auto letterbox, 16:9 → 9:16 vertical via crop+scale.
Step 8 — Speed / time manipulation
ffmpeg-speed-time: setpts, atempo, minterpolate for smooth slow-mo, freeze frame, reverse.
Step 9 — Chromakey / greenscreen
ffmpeg-chromakey: chromakey + despill filters.
Step 10 — Text / lower thirds
drawtext or ASS sidecar subtitles.
Step 11 — Subtitles
ffmpeg-subtitles: soft-mux, burn-in, extract, format convert.
Step 12 — Audio finishing
EQ + compression via ffmpeg-audio-filter, loudness via media-ffmpeg-normalize.
Step 13 — Thumbnails / sprite sheets
ffmpeg-frames-images: extract every N seconds, sprite sheet for scrubber UI, animated GIF preview.
Step 14 — Chapters + metadata
ffmpeg-metadata: chapter markers + cover art.
Step 15 — Final delivery
HandBrake preset for prosumer; ffmpeg direct for fine-grained; AV1 for archival/bandwidth.
Step 16 — QC
workflow-analysis-quality — VMAF vs reference, spec assertions.
Step 17 — Upload
media-cloud-upload (YouTube, Vimeo, S3, etc.).
Variants
- Batch —
media-batch runs finish-xxx.sh across N files in parallel.
- Programmatic (MoviePy) — when ffmpeg filtergraph gets unwieldy, drop to Python.
- Vertical social — face-tracking auto-crop 16:9 → 9:16.
- 360° VR edit — equirect → stereographic "little planet" via
ffmpeg-360-3d.
Gotchas
-movflags +faststart is a SECOND pass. Encoder runs; ffmpeg rewinds; rewrites moov atom to front. Adds time — necessary for progressive MP4.concat demuxer requires identical codecs. Different rates / codecs = silent failure. Use concat FILTER instead.- Slow-motion
setpts=N*PTS needs matching -r input. Without, frame drop/duplicate is odd.
atempo range is 0.5–2.0 per filter. Chain for extreme tempo shifts.yadif modes: 0=single-rate, 1=double-rate, 2=spatial-only, 3=double-rate spatial.scale=-2:1080 keeps aspect ratio with even width (required for yuv420p).libx264 -pix_fmt yuv420p required for broadcast playback. Many decoders don't handle yuv422p / yuv444p.- NVENC preset is
p1–p7 (higher = better quality, slower). NOT ultrafast–placebo.
h264_videotoolbox is quality-preset only. -b:v and -q:v are hints; quality varies.libsvtav1 -preset is 0–13 (0 slowest). NOT 1–10 like libaom-av1.libvpx-vp9 needs -row-mt 1 -threads 4 to multi-thread. Single-threaded default.-shortest stops at shortest input. Without it, looping -loop 1 inputs = infinite.drawtext needs font FILE PATH, not font name. macOS /System/Library/Fonts/Helvetica.ttc.- ASS subtitles use
&HAABBGGRR (ARGB). AA=alpha, BGR byte order. Wrong byte order = wrong color.
chromakey works in RGB. YUV sources auto-convert, but precise keying wants format=yuva420p first.vidstab requires --enable-libvidstab ffmpeg build. Some Ubuntu stock builds lack it.atempo preserves pitch; asetrate shifts pitch. Combine carefully.-ss before -i = fast seek, non-frame-accurate. -ss after -i = frame-accurate, sequential. Use after for accuracy.- Two-pass loudnorm via
media-ffmpeg-normalize. Single-pass applies runtime auto limits.
-map 0:v:0 -map 0:a:0 selects first video+audio. Default -map 0 grabs ALL streams.- HandBrake presets are opinionated. "Fast 1080p30" = H.264 CRF 22 + AAC 160. Sometimes wrong for source.
- Alpha preservation needs
-pix_fmt yuva420p (or yuva444p) AND -c:v qtrle or -c:v prores_ks -profile:v 4 (ProRes 4444). H.264 has NO alpha.
Example — 4K camera raw → 1080p YouTube master
Probe → ProRes 422 HQ mezzanine → cut + concat → LUT color grade → hqdn3d → scale=-2:1080 → loudnorm two-pass → drawtext lower third → mov_text subtitles → H.264 CRF 18 yuv420p -movflags +faststart → VMAF QC → YouTube upload.
Related
workflow-editorial-interchange — round-tripping the edit into/out of NLEs.workflow-ai-enhancement — AI upscale/interpolate/denoise path.workflow-analysis-quality — QC gating before shipping.
