# Demo gallery Every clip on this page but three was generated from nothing: ffmpeg's own synthetic sources make the footage. The three exceptions, `reframe_talk`, `filler_en` and `pauses_en`, use AI-generated footage of people who do not exist (`demos/footage/`). Each "after" is produced by running one of this repo's scripts on its source. Rebuild the whole page's material with: ```bash python3 demos/build.py # or: npm run demo python3 demos/build.py --list # what gets built python3 demos/build.py --only captions_ja ``` The full-resolution `_before.mp4`, `_after.mp4` and side-by-side `.mp4` land in `demos/out/` (gitignored). Only the small previews below are committed, and the build fails if any of them exceeds 500 KB. In every preview the left half is the input and the right half is what the command produced -- except where the burnt-in labels say otherwise, for a demo comparing two settings (a lossless cut against an accurate one) rather than a before against an after. The tools at the bottom, under **Inspection**, have no picture: they answer a question instead of changing one, so the command is the demo. ## Captions & text ### Burned-in captions (English) ![Burned-in captions (English)](demos/captions_en.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_en.txt --size 30 --bold --position bottom --margin 40 -o demos/out/captions_en_after.mp4 ``` **Look for:** Plain-text cues become an SRT and are rendered by libass -- outline and margin come from the flags, not from a template. ### Phrase-aware line breaks ![Phrase-aware line breaks](demos/captions_phrase.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_phrase.txt --size 34 --bold --position bottom --margin 40 --wrap measured --preset veryfast -o demos/out/captions_phrase_before.mp4 python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_phrase.txt --size 34 --bold --position bottom --margin 40 --wrap phrase --preset veryfast -o demos/out/captions_phrase_after.mp4 ``` **Look for:** The same cue at the same size, with one flag different. Left is 1.15's wrap, which only minimised the widest line and left `the` and `con` stranded at the end of a line; right is 1.16's default, which never breaks inside a word and never ends a line on an article or a preposition. The text itself is untouched -- the skill never rewrites a caption to make it fit. ### Caption size fitted to the cue ![Caption size fitted to the cue](demos/captions_fitsize.gif) ```bash python3 scripts/caption.py demos/out/fixtures/vertical.mp4 --text demos/out/fixtures/cues_phrase.txt --platform tiktok --max-lines 2 --fit-size off --bold --preset veryfast -o demos/out/captions_fitsize_before.mp4 python3 scripts/caption.py demos/out/fixtures/vertical.mp4 --text demos/out/fixtures/cues_phrase.txt --platform tiktok --max-lines 2 --fit-size on --bold --preset veryfast -o demos/out/captions_fitsize_after.mp4 ``` **Look for:** The same cue file at the same destination, with one flag different. Left is 1.16: at the TikTok caption size the cue cannot fit two lines, so it is split into two consecutive cues and half the sentence arrives late. Right is 1.17's default: the size dropped until the whole sentence is on screen at once, and stopped well above the 4.5 %-of-frame-height floor. The text is untouched -- the skill never rewrites a caption to make it fit. ### Three subtitle tracks in one file ![Three subtitle tracks in one file](demos/captions_multitrack.gif) ```bash python3 scripts/caption.py demos/out/fixtures/mandel.mp4 --mode mux --srt demos/out/captions_multitrack_track_en.srt:en --srt demos/out/captions_multitrack_track_ja.srt:ja --srt demos/out/captions_multitrack_track_es.srt:es --default-track en -o demos/out/captions_multitrack_muxed.mkv ``` **Look for:** Nothing is burnt in and nothing is re-encoded: the video and audio are copied bit for bit and three SRTs ride along as toggleable, language-tagged streams. The right half is the track list ffprobe reads back. ### Japanese captions ![Japanese captions](demos/captions_ja.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_ja.txt --size 30 --bold --position bottom --margin 40 -o demos/out/captions_ja_after.mp4 ``` **Look for:** The font is chosen per script: Japanese cues get a CJK face automatically, so no box-glyph tofu appears. ### Chinese captions ![Chinese captions](demos/captions_zh.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_zh.txt --size 30 --bold --position bottom --margin 40 -o demos/out/captions_zh_after.mp4 ``` **Look for:** Same command, Han text: line breaking and the font switch are handled without a --font flag. ### Korean captions ![Korean captions](demos/captions_ko.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_ko.txt --size 30 --bold --position bottom --margin 40 -o demos/out/captions_ko_after.mp4 ``` **Look for:** Hangul wins script detection even when Latin words are mixed into the same cue. ### Arabic captions ![Arabic captions](demos/captions_ar.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_ar.txt --size 30 --bold --position bottom --margin 40 -o demos/out/captions_ar_after.mp4 ``` **Look for:** Right-to-left text shaped by libass; the Latin fragments inside it stay left-to-right. ### Animated pop captions with karaoke ![Animated pop captions with karaoke](demos/captions_pop_karaoke.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_pop.txt --size 30 --bold --position bottom --margin 40 --animate pop --karaoke --highlight-color '#39ff88' -o demos/out/captions_pop_karaoke_after.mp4 ``` **Look for:** Each cue scales in, and the highlight colour walks word by word across the line. ### Emoji captions in colour ![Emoji captions in colour](demos/captions_emoji.gif) ```bash python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_emoji.txt --size 30 --bold --margin 40 --emoji mono -o demos/out/captions_emoji_before.mp4 python3 scripts/caption.py demos/out/fixtures/motion.mp4 --text demos/out/fixtures/cues_emoji.txt --size 30 --bold --margin 40 --emoji-assets demos/out/fixtures/emoji -o demos/out/captions_emoji_after.mp4 ``` **Look for:** Left: emoji as this ffmpeg's libass draws them (monochrome, or missing). Right: one PNG per emoji cluster composited over the caption, with the ASS reserving the exact gap -- the text does not move. ### Hindi lower third, shaped ![Hindi lower third, shaped](demos/lower_third_hindi.gif) ```bash python3 scripts/graphics.py demos/out/fixtures/motion.mp4 --template lower-third --name 'प्रिया शर्मा' --title 'निर्देशक' --start 0.5 --end 5 --text-render ass -o demos/out/lower_third_hindi_after.mp4 ``` **Look for:** Left: the same text through drawtext -- the i-matra is not reordered and the final matra is dropped. Right: graphics.py routing Devanagari through libass automatically. drawtext does not use harfbuzz on any build. ### Lower third ![Lower third](demos/lower_third.gif) ```bash python3 scripts/graphics.py demos/out/fixtures/motion.mp4 --template lower-third --name 'Ada Lovelace' --title 'Analytical Engine' --start 0.5 --end 5 -o demos/out/lower_third_after.mp4 ``` **Look for:** Name and role slide in from the left over the picture and slide out again -- no image asset involved. ### Title card ![Title card](demos/title_card.gif) ```bash python3 scripts/graphics.py demos/out/fixtures/mandel.mp4 --template title --title 'Episode 12' --subtitle 'The math of video' --start 0 --end 4 -o demos/out/title_card_after.mp4 ``` **Look for:** A centred title and subtitle fade in over the first seconds and leave the rest of the clip untouched. ## Picture ### Logo overlay with fade ![Logo overlay with fade](demos/logo_overlay.gif) ```bash python3 scripts/overlay.py demos/out/fixtures/motion.mp4 --image demos/out/fixtures/logo.png --position top-right --scale 200 --opacity 0.9 --start 0.5 --end 7.5 --fade 0.6 -o demos/out/logo_overlay_after.mp4 ``` **Look for:** The semi-transparent logo fades in at 0.5 s and out before the end; the underlying picture is unchanged. ### HDR10 to SDR ![HDR10 to SDR](demos/hdr_tonemap.gif) ```bash python3 scripts/color.py demos/out/fixtures/hdr10.mp4 --to-sdr --tonemap hable --preset veryfast -o demos/out/hdr_tonemap_after.mp4 ``` **Look for:** The PQ / BT.2020 source is tone-mapped to BT.709: on an SDR screen the 'before' is the washed-out one. ### 16:9 to 9:16 by cropping ![16:9 to 9:16 by cropping](demos/reframe_crop.gif) ```bash python3 scripts/fit.py demos/out/fixtures/mandel.mp4 --aspect 9:16 --fit crop --width 540 --preset veryfast -o demos/out/reframe_crop_after.mp4 ``` **Look for:** The vertical frame is cut out of the centre of the wide one -- full height, sides lost. ### A talking head made vertical (AI-generated footage) ![A talking head made vertical (AI-generated footage)](demos/reframe_talk.gif) ```bash python3 scripts/fit.py demos/footage/ai_talk.mp4 --aspect 9:16 --fit crop --crop-x 0.27 --width 540 --preset veryfast -o demos/out/reframe_talk_after.mp4 ``` **Look for:** The 9:16 frame is placed on the speaker with --crop-x 0.27 instead of the centre, where it would cut her face in half. The source is AI-generated footage, not a real person. ### 16:9 to 9:16 by padding ![16:9 to 9:16 by padding](demos/reframe_pad.gif) ```bash python3 scripts/fit.py demos/out/fixtures/mandel.mp4 --aspect 9:16 --fit pad --pad-fill blur --width 540 --preset veryfast -o demos/out/reframe_pad_after.mp4 ``` **Look for:** Nothing is lost: the wide frame is kept whole and the gap above and below is filled with a blurred copy. ### Speed change to hit a duration ![Speed change to hit a duration](demos/speed_up.gif) ```bash python3 scripts/fit.py demos/out/fixtures/motion.mp4 --duration 4 --method speed --preset veryfast -o demos/out/speed_up_after.mp4 ``` **Look for:** An 8 s clip retimed to land exactly on 4 s; audio is pitch-corrected rather than chipmunked. ### Reverse ![Reverse](demos/reverse.gif) ```bash python3 scripts/reverse.py demos/out/fixtures/life.mp4 --preset veryfast -o demos/out/reverse_after.mp4 ``` **Look for:** The life pattern runs backwards -- cells un-die; the audio is reversed with it. ### Join with a cross fade ![Join with a cross fade](demos/join_fade.gif) ```bash python3 scripts/join.py demos/out/fixtures/motion.mp4 demos/out/fixtures/mandel.mp4 --transition fade --duration 0.8 --width 960 --preset veryfast -o demos/out/join_fade_after.mp4 ``` **Look for:** Two clips of different content become one; watch the 0.8 s dissolve in the middle. ### Join through black ![Join through black](demos/join_fadeblack.gif) ```bash python3 scripts/join.py demos/out/fixtures/motion.mp4 demos/out/fixtures/mandel.mp4 --transition fadeblack --duration 0.8 --width 960 --preset veryfast -o demos/out/join_fadeblack_after.mp4 ``` **Look for:** The same join with fadeblack: the cut dips to black instead of blending the two pictures. ### 16:9 to 9:16 on a blurred background ![16:9 to 9:16 on a blurred background](demos/fit_blur.gif) ```bash python3 scripts/fit.py demos/out/fixtures/mandel.mp4 --aspect 9:16 --fit blur --width 540 --preset veryfast -o demos/out/fit_blur_after.mp4 ``` **Look for:** Nothing is cropped and there are no black bars: the whole wide frame sits centred on a blurred, dimmed copy of itself. ### Lossless cut vs. accurate cut ![Lossless cut vs. accurate cut](demos/cut_accurate.gif) ```bash python3 scripts/cut.py demos/out/fixtures/mandel.mp4 --start 2.05 --duration 3 --tolerance 5 --json -o demos/out/cut_accurate_before.mp4 python3 scripts/cut.py demos/out/fixtures/mandel.mp4 --start 2.05 --duration 3 --accurate --preset veryfast -o demos/out/cut_accurate_after.mp4 ``` **Look for:** Both sides asked for the same 2.05 s start. The stream copy could only snap to the nearest keyframe, so its first frame is from earlier in the clip; --accurate re-encodes and starts on the frame that was asked for. ### Cut out a range, keep the rest ![Cut out a range, keep the rest](demos/cut_remove.gif) ```bash python3 scripts/cut.py demos/out/fixtures/mandel.mp4 --remove 2-4 --accurate --preset veryfast --json -o demos/out/cut_remove_after.mp4 ``` **Look for:** cut.py --remove 2-4 drops two seconds from the middle and joins what is left: the 'after' side jumps from 2 s to 4 s of the source and finishes two seconds early. ### Frame-exact cut, re-encoding only the ends ![Frame-exact cut, re-encoding only the ends](demos/cut_smart.gif) ```bash python3 scripts/cut.py demos/out/cut_smart_gop30.mp4 --start 1.3 --end 6.6 --smart --json -o demos/out/cut_smart_after.mp4 ``` **Look for:** cut.py --smart cuts 1.3-6.6 s frame-exactly but re-encodes only 1.3-2 s and 6-6.6 s, matched to the source's H.264 parameters; the whole GOPs between are the source's own packets, checked against it before the file is written. The build fails unless the result is a verified smart cut. ### Lossless cut that starts where asked ![Lossless cut that starts where asked](demos/cut_edit_list.gif) ```bash python3 scripts/cut.py demos/out/fixtures/mandel.mp4 --start 2.05 --duration 3 --tolerance 5 --json -o demos/out/cut_edit_list_before.mp4 python3 scripts/cut.py demos/out/fixtures/mandel.mp4 --start 2.05 --duration 3 --tolerance 5 --edit-list --json -o demos/out/cut_edit_list_after.mp4 ``` **Look for:** Both sides are stream copies of the same 2.05 s request. The default shifts the copy's timestamps to zero, so the keyframe's pre-roll plays first (start_snapped: true); --edit-list keeps the MP4 edit list that hides it, so the picture starts at 2.05 s (start_snapped: false, stored_preroll_seconds says how much is hidden). A player that ignores edit lists shows the pre-roll. ### Cuts that land on the beat ![Cuts that land on the beat](demos/beats_snap.gif) ```bash python3 scripts/cut.py demos/out/fixtures/clicks.mp4 --start 2.03 --end 6.01 --snap none --accurate --preset veryfast -o demos/out/beats_snap_before.mp4 python3 scripts/cut.py demos/out/fixtures/clicks.mp4 --start 2.03 --end 6.01 --snap beats --accurate --preset veryfast -o demos/out/beats_snap_after.mp4 ``` **Look for:** Both cuts asked for 2.03 s. The right one moved 30 ms to the nearest measured onset, so its first frame lands on a click instead of just after one; the waveform is the evidence. The tempo, the beat list and the confidence are in the JSON -- nothing is snapped to a grid the audio does not support, and below --min-confidence the cut refuses rather than inventing one. ### Crop to an exact rectangle ![Crop to an exact rectangle](demos/crop_rect.gif) ```bash python3 scripts/crop.py demos/out/fixtures/mandel.mp4 --x 320 --y 120 --width 640 --height 480 --preset veryfast -o demos/out/crop_rect_after.mp4 ``` **Look for:** A literal 640x480 window at x=320, y=120 in the source frame -- no aspect maths, no auto-centring; the rectangle is the caller's and is refused rather than rounded if it does not fit. ### Detect letterbox bars, then remove them ![Detect letterbox bars, then remove them](demos/cropdetect_crop.gif) ```bash python3 scripts/cropdetect.py demos/out/fixtures/letterbox.mp4 --seconds 4 --json python3 scripts/crop.py demos/out/fixtures/letterbox.mp4 --x 0 --y 160 --width 1280 --height 400 --preset veryfast -o demos/out/cropdetect_crop_after.mp4 ``` **Look for:** cropdetect.py measures the bars and prints the rectangle; crop.py is what actually cuts them off. Measuring and deciding stay two commands. ### Pad the timeline with black ![Pad the timeline with black](demos/pad_timeline.gif) ```bash python3 scripts/pad.py demos/out/fixtures/mandel.mp4 --start 1 --end 1 --color 0x101014 --preset veryfast -o demos/out/pad_timeline_after.mp4 ``` **Look for:** A second of black and silence before and after the clip -- the TIMELINE grows, the frame does not (that is fit.py --fit pad). ### Freeze frame under a title ![Freeze frame under a title](demos/freeze_hold.gif) ```bash python3 scripts/graphics.py demos/out/fixtures/mandel.mp4 --template title --title 'Hold this frame' --subtitle 'freeze.py --at 2 --hold 1.5' --start 1.6 --end 4 -o demos/out/freeze_hold_titled.mp4 python3 scripts/freeze.py demos/out/freeze_hold_titled.mp4 --at 2 --hold 1.5 --preset veryfast -o demos/out/freeze_hold_after.mp4 ``` **Look for:** The clip stops dead at 2 s for a beat and a half while the title card sits over it, then carries on -- everything after the freeze is pushed later, nothing is lost. ### Loop a short clip ![Loop a short clip](demos/loop_times.gif) ```bash python3 scripts/cut.py demos/out/fixtures/life.mp4 --start 0 --duration 1.5 --accurate --preset veryfast -o demos/out/loop_times_before.mp4 python3 scripts/loop.py demos/out/loop_times_before.mp4 --times 4 --preset veryfast -o demos/out/loop_times_after.mp4 ``` **Look for:** A 1.5 s clip repeated four times back to back. The seam is not smoothed: a clip that does not already loop cleanly shows its cut (--crossfade, below, fades it). ### A seamless loop ![A seamless loop](demos/loop_crossfade.gif) ```bash python3 scripts/cut.py demos/out/fixtures/life.mp4 --start 0 --duration 1.5 --accurate --preset veryfast -o demos/out/loop_crossfade_clip.mp4 python3 scripts/loop.py demos/out/loop_crossfade_clip.mp4 --times 4 --preset veryfast -o demos/out/loop_crossfade_before.mp4 python3 scripts/loop.py demos/out/loop_crossfade_clip.mp4 --times 4 --crossfade 0.4 --preset veryfast -o demos/out/loop_crossfade_after.mp4 ``` **Look for:** The same clip looped four times, plain (left) and with loop.py --crossfade 0.4 (right): the plain loop jumps back at every seam, the crossfaded one fades the clip's last 0.4 s into its first, so each repeat runs on into the next. ### A still card dropped into the timeline ![A still card dropped into the timeline](demos/insert_still.gif) ```bash python3 scripts/insert.py demos/out/fixtures/slate.png --duration 2 --width 960 --height 540 --fps 25 --zoom in --pan right --preset veryfast -o demos/out/insert_still_card.mp4 python3 scripts/join.py demos/out/fixtures/motion.mp4 demos/out/insert_still_card.mp4 demos/out/fixtures/mandel.mp4 --transition fade --duration 0.5 --width 960 --height 540 --fps 25 --preset veryfast -o demos/out/insert_still_after.mp4 ``` **Look for:** insert.py turns one PNG into a 2 s clip with a slow zoom and pan, and join.py cross-fades it in between the two shots. ### Numbered stills into a clip ![Numbered stills into a clip](demos/sequence_frames.gif) ```bash python3 scripts/sequence.py --dir demos/out/fixtures/frames --pattern frame_%04d.png --start-number 1 --fps 8 --preset veryfast -o demos/out/sequence_frames_after.mp4 ``` **Look for:** Twenty-four PNGs named frame_0001.png onwards become one 8 fps clip; the frame list is resolved and checked on disk before ffmpeg is asked to read it. ### Speed ramp ![Speed ramp](demos/speedramp.gif) ```bash python3 scripts/speedramp.py demos/out/fixtures/mandel.mp4 --segment 0-2:2.0 --segment 2-3:0.35 --segment 3-6:1.5 --preset veryfast -o demos/out/speedramp_after.mp4 ``` **Look for:** Three constant-speed segments in one pass: 2x, then a 0.35x slam for the beat, then 1.5x out. Audio is pitch-corrected through all three. ### Denoise ![Denoise](demos/denoise.gif) ```bash python3 scripts/denoise.py demos/out/fixtures/noisy.mp4 --strength high --preset veryfast -o demos/out/denoise_after.mp4 ``` **Look for:** hqdn3d at --strength high takes the grain out; look closely and the fine detail softens with it, which is the trade the flag is making. ### Deinterlace ![Deinterlace](demos/deinterlace.gif) ```bash python3 scripts/deinterlace.py demos/out/fixtures/interlaced.mp4 --mode frame --parity tff --preset veryfast -o demos/out/deinterlace_after.mp4 ``` **Look for:** The combing on moving edges -- alternate lines from two different moments -- is woven back into whole progressive frames by yadif. ### Stabilize shaky footage ![Stabilize shaky footage](demos/stabilize.gif) ```bash python3 scripts/stabilize.py demos/out/fixtures/shaky.mp4 --shakiness 8 --smoothing 20 --zoom 5 --preset veryfast -o demos/out/stabilize_after.mp4 ``` **Look for:** vidstab's two passes cancel the handheld jitter and crop in 5% to hide the edges that smoothing exposes; the frame stops wandering. ### Straighten a tilted horizon ![Straighten a tilted horizon](demos/straighten.gif) ```bash python3 scripts/straighten.py demos/out/fixtures/tilted.mp4 --degrees -5 --fit crop --preset veryfast -o demos/out/straighten_after.mp4 ``` **Look for:** A 5 degree tilt taken back out. --fit crop scales up just enough that the rotated corners leave no gap, costing a thin border of picture. ### Redact a rectangle ![Redact a rectangle](demos/redact_blur.gif) ```bash python3 scripts/redact.py demos/out/fixtures/motion.mp4 --x 420 --y 180 --width 440 --height 360 --mode blur --blur-strength 24 --preset veryfast -o demos/out/redact_blur_after.mp4 ``` **Look for:** One region blurred for the whole clip and the rest of the frame untouched. The rectangle has to be given: this tool finds no faces and no plates. ### Remove a fixed logo ![Remove a fixed logo](demos/delogo.gif) ```bash python3 scripts/redact.py demos/out/delogo_before.mp4 --x 48 --y 36 --width 240 --height 96 --mode delogo --preset veryfast -o demos/out/delogo_after.mp4 ``` **Look for:** The white block is a stand-in for a burned-in logo. redact.py --mode delogo fills that rectangle from the picture around it. A logo that moves, or text that needs inpainting, is out of scope: the box is the caller's, and it has to sit 1 px inside the frame. ### 360 panorama to a flat view ![360 panorama to a flat view](demos/sphere_flat.gif) ```bash python3 scripts/sphere.py demos/out/fixtures/equirect.mp4 --input-projection equirect --yaw 40 --pitch -10 --h-fov 100 --v-fov 70 --width 960 --height 540 --preset veryfast -o demos/out/sphere_flat_after.mp4 ``` **Look for:** v360 maps the 2:1 equirectangular source onto a rectilinear camera pointed 40 degrees right and 10 degrees down -- the same viewport a headset would show, baked into a file. ### Green screen onto a generated background ![Green screen onto a generated background](demos/greenscreen.gif) ```bash python3 scripts/background.py --duration 4 --width 960 --height 540 --gradient 0xff6a00:0x0057ff --angle 45 --fps 25 --preset veryfast -o demos/out/greenscreen_plate.mp4 python3 scripts/overlay.py demos/out/greenscreen_plate.mp4 --video demos/out/fixtures/green.mp4 --chromakey 0x00b140 --chromakey-similarity 0.18 --chromakey-blend 0.05 --position center --scale 960 --preset veryfast -o demos/out/greenscreen_after.mp4 ``` **Look for:** background.py draws the gradient plate from nothing (there is no input file), then overlay.py keys the chroma green out of the subject and composites it on top. ### A graphic rendered as its own alpha layer ![A graphic rendered as its own alpha layer](demos/alpha_layer.gif) ```bash python3 scripts/background.py --transparent --alpha-codec qtrle --duration 6 --width 1280 --height 720 --fps 30 -o demos/out/alpha_layer_canvas.mov --json python3 scripts/graphics.py demos/out/alpha_layer_canvas.mov --template lower-third --name 'Ada Lovelace' --title 'Analytical Engine' --start 0.5 --end 5.5 --keep-alpha --alpha-codec qtrle -o demos/out/alpha_layer_layer.mov --json python3 scripts/overlay.py demos/out/fixtures/motion.mp4 --video demos/out/alpha_layer_layer.mov --position 0,0 --margin 0 --preset veryfast -o demos/out/alpha_layer_after.mp4 ``` **Look for:** background.py --transparent makes an empty canvas, graphics.py --keep-alpha draws a lower third on it and writes it with its alpha (QuickTime Animation here; ProRes 4444 by default), checked frame by frame, and overlay.py lays that one layer over the footage. The layer can go to an editor as it is: nothing but the graphic was ever in its picture. ### Pieces of one layer at their own times ![Pieces of one layer at their own times](demos/overlay_place.gif) ```bash python3 scripts/overlay.py demos/out/fixtures/motion.mp4 --video demos/out/fixtures/mandel.mp4 --position bottom-right --scale 480 --place 0-1.5@0.5 --place 4-5.5@3.5 --preset veryfast -o demos/out/overlay_place_after.mp4 ``` **Look for:** overlay.py --place lays two ranges of the fractal clip on the test pattern: its first 1.5 s at 0.5 s, its 4-5.5 s at 3.5 s. Between them the layer is hidden and the main picture is untouched -- one pass, no full-length mostly-transparent layer to render first. ### Apply a .cube LUT ![Apply a .cube LUT](demos/color_lut.gif) ```bash python3 scripts/color.py demos/out/fixtures/motion.mp4 --lut demos/out/fixtures/look.cube --lut-strength 1.0 --preset veryfast -o demos/out/color_lut_after.mp4 ``` **Look for:** A 3D LUT applied at full strength: highlights warm, shadows cool. The same flag takes a camera vendor's Log-to-709 transform or a creative look -- the file decides, the tool does not. ### Primary colour correction ![Primary colour correction](demos/color_correct.gif) ```bash python3 scripts/color.py demos/out/fixtures/motion.mp4 --correct --exposure 0.25 --contrast 1.25 --saturation 1.3 --temperature 7000 --gamma 1.1 --preset veryfast -o demos/out/color_correct_after.mp4 ``` **Look for:** Typed values, not a look: exposure +0.25, contrast 1.25, saturation 1.3, white balance to 7000 K, gamma 1.1 -- each one a number the caller chose. ### Broadcast legal levels ![Broadcast legal levels](demos/color_legalize.gif) ```bash python3 scripts/color.py demos/out/color_legalize_hot.mp4 --correct --legalize bt709 --quality 8 --preset veryfast -o demos/out/color_legalize_legal.mp4 --json ``` **Look for:** A grade that runs hot, before and after color.py --legalize bt709: each half shows the picture over its luma waveform, whose green graticule marks the digital scale. The clamp acts before encoding, on the 4:2:0 picture the encoder receives (16-235 luma, 16-240 chroma); the encoder's error at the clamped edges can put a few decoded samples back out. The run then measures the encoded file (check.py --levels counts the same way), which passes with up to 1 % of a frame's luma or chroma samples out of range, and the build fails unless that check passes; --quality 8 keeps this pattern inside the allowance. ### 2x2 comparison grid ![2x2 comparison grid](demos/grid_2x2.gif) ```bash python3 scripts/grid.py demos/out/fixtures/motion.mp4 demos/out/fixtures/mandel.mp4 demos/out/fixtures/life.mp4 demos/out/fixtures/camb.mp4 --cols 2 --rows 2 --cell-width 480 --cell-height 270 --fps 25 --audio-from 0 --gap 4 --preset veryfast -o demos/out/grid_2x2_after.mp4 ``` **Look for:** Four clips of different content in one frame, each letterboxed into its cell rather than stretched, with the source name burnt into the corner and only input 0's audio carried through. ### B-roll cutaway ![B-roll cutaway](demos/broll_cutaway.gif) ```bash python3 scripts/broll.py demos/out/fixtures/motion.mp4 --insert demos/out/fixtures/mandel.mp4 --at 2 --duration 3 --audio a --preset veryfast -o demos/out/broll_cutaway_after.mp4 ``` **Look for:** From 2 s to 5 s the picture cuts away to the insert while the interview audio keeps running underneath, then comes back at its own time -- the output is exactly as long as the A-roll. ### Two cameras, one cut list ![Two cameras, one cut list](demos/multicam_switch.gif) ```bash python3 scripts/multicam.py demos/out/fixtures/motion.mp4 demos/out/fixtures/camb.mp4 --switch 0-2:0,2-4:1,4-6:0 --width 960 --height 540 --fps 25 --preset veryfast -o demos/out/multicam_switch_after.mp4 ``` **Look for:** Camera B's audio is the same event 0.6 s offset; multicam.py measures that by cross-correlation, aligns both, then cuts between them on the ranges it was given. ### Crop centred on the measured motion ![Crop centred on the measured motion](demos/motion_centre_crop.gif) ```bash python3 scripts/cropdetect.py demos/out/motion_centre_crop_offcentre.mp4 --motion-centre --json python3 scripts/crop.py demos/out/motion_centre_crop_offcentre.mp4 --x 328 --y 0 --width 304 --height 540 --preset veryfast -o demos/out/motion_centre_crop_centre.mp4 python3 scripts/crop.py demos/out/motion_centre_crop_offcentre.mp4 --x 632 --y 0 --width 304 --height 540 --preset veryfast -o demos/out/motion_centre_crop_after.mp4 ``` **Look for:** cropdetect.py --motion-centre reports where the motion sits per second; it never picks a reframe itself. Left is a naive centre crop that clips the moving subject sitting off to one side; right is crop.py aimed at the measured centroid instead. ### Auto-switch to whichever camera is loudest ![Auto-switch to whichever camera is loudest](demos/multicam_switch_energy.gif) ```bash python3 scripts/multicam.py demos/out/fixtures/motion.mp4 demos/out/fixtures/camb.mp4 --switch 0-6:0 --width 960 --height 540 --fps 25 --preset veryfast -o demos/out/multicam_switch_energy_single_cam.mp4 python3 scripts/multicam.py demos/out/fixtures/motion.mp4 demos/out/fixtures/camb.mp4 --switch energy --min-shot 1 --width 960 --height 540 --fps 25 --preset veryfast -o demos/out/multicam_switch_energy_after.mp4 ``` **Look for:** The literal 'energy' switch list auto-cuts to whichever of the two cameras is loudest at each moment, respecting --min-shot; left is a fixed single-camera cut for contrast. ## Audio ### Filler words cut from real speech (AI-generated footage) ![Filler words cut from real speech (AI-generated footage)](demos/filler_en.gif) ```bash python3 scripts/silence.py demos/out/filler_en_talk.mp4 --filler --words demos/footage/ai_talk_en.words.json --margin 0 --preset veryfast -o demos/out/filler_en_after.mp4 ``` **Look for:** "Hey, er, hello there. Well, um, good morning." Right is silence.py --filler: the er and the um are gone and the sentence plays straight through. The source is AI-generated footage, not a real person. ### Pauses tightened, more room left after a sentence (AI-generated footage) ![Pauses tightened, more room left after a sentence (AI-generated footage)](demos/pauses_en.gif) ```bash python3 scripts/silence.py demos/out/pauses_en_talk.mp4 --words demos/footage/ai_talk_en.words.json --threshold -30 --min-silence 0.2 --sentence-min-silence 0.3 --margin-lead 0.1 --margin-tail 0.15 --sentence-margin-tail 0.3 --preset veryfast -o demos/out/pauses_en_after.mp4 ``` **Look for:** "Hey, er, hello there. Well, um, good morning. Hope you're having a good day." Right is silence.py with --sentence-min-silence and --sentence-margin-tail: the 1.5 s pauses after each sentence shrink to 0.4 s (0.3 s kept after the sentence, 0.1 s before the next), a pause inside a sentence would keep 0.15 s. The source is AI-generated footage, not a real person. ### Silence removal ![Silence removal](demos/silence_removal.gif) ```bash python3 scripts/silence.py demos/out/fixtures/motion.mp4 --threshold -35 --min-silence 0.4 --margin 0.1 --timemap demos/out/silence_removal_timemap.csv --preset veryfast -o demos/out/silence_removal_after.mp4 ``` **Look for:** The 'after' side runs out of material and freezes: that held frame is the part of the timeline that was cut. The --timemap CSV beside it says where each kept range sits in the source and in the cut, to re-time subtitles. ### Loudness normalisation to -14 LUFS ![Loudness normalisation to -14 LUFS](demos/loudness.gif) ```bash python3 scripts/loudness.py demos/out/loudness_before.mp4 -I -14 --tp -1 -o demos/out/loudness_after.mp4 ``` **Look for:** Two showwavespic plots: the quiet input on the left, the same programme brought up to broadcast level on the right without clipping. ### 5.1 to stereo downmix ![5.1 to stereo downmix](demos/downmix_51.gif) ```bash python3 scripts/audio.py demos/out/fixtures/surround.mp4 --downmix -o demos/out/downmix_51_after.mp4 ``` **Look for:** Six discrete tones folded into two channels with the ITU-R BS.775 weights -- the centre and surrounds are still audible, the LFE is dropped, and nothing clips. ### Music bed with ducking ![Music bed with ducking](demos/bgm_ducking.gif) ```bash python3 scripts/audio.py demos/out/fixtures/motion.mp4 --music demos/out/fixtures/music.m4a --duck --duck-amount 12 --music-volume 0.6 --fade-out 1.5 -o demos/out/bgm_ducking_after.mp4 ``` **Look for:** The bed drops by 12 dB whenever the speech-like track is active and comes back up in the pauses. ### Faster at the same pitch, to an exact length ![Faster at the same pitch, to an exact length](demos/tempo_exact_length.gif) ```bash python3 scripts/audio.py demos/out/fixtures/music.m4a --tempo 1.25 --sample-rate 44100 --length-samples 176400 -o demos/out/tempo_exact_length_after.wav ``` **Look for:** The bed plays 1.25x faster without its pitch moving, resampled to 44.1 kHz, and is cut at sample 176400 (4 s): the right-hand plot holds more of the bed's pulses in half the time. The result's length_samples step counts the written samples. ### External mic aligned to the camera ![External mic aligned to the camera](demos/sync_mic.gif) ```bash python3 scripts/sync.py demos/out/fixtures/motion.mp4 demos/out/fixtures/mic.wav --replace-audio --json -o demos/out/sync_mic_after.mp4 ``` **Look for:** The mic was started after the camera, so its envelope begins late in its own file; after the sync the same speech sits where the camera's does. Left is the mic as recorded, right is the track that ends up on the picture -- the measured offset and its confidence are printed by the command. ### Audio rendered as a waveform video ![Audio rendered as a waveform video](demos/waveform_render.gif) ```bash python3 scripts/waveform.py demos/out/fixtures/music.m4a --style waveform --width 960 --height 540 --color 0x39ff88 --background 0x101014 --waveform-mode cline --preset veryfast -o demos/out/waveform_render_after.mp4 ``` **Look for:** An audio-only file has nothing to show; showwaves draws the amplitude as it plays, and the rendered clip carries the same audio it is drawing. ### Audio episode as a shareable audiogram ![Audio episode as a shareable audiogram](demos/audiogram.gif) ```bash python3 scripts/background.py --width 960 --height 540 --gradient '#101014:#2a1846' --duration 1 -o demos/out/audiogram_plate.mp4 python3 scripts/waveform.py demos/out/fixtures/music.m4a --image demos/out/audiogram_plate.png --width 960 --height 540 --position strip --vis-height 0.3 --color 0x39ff88 --title 'Episode 12' --text demos/out/audiogram_cues.txt --preset veryfast -o demos/out/audiogram_after.mp4 ``` **Look for:** The same track as the waveform demo, this time over a still plate with the episode title and burnt-in captions -- one waveform.py --image call (or render.py --template audiogram). The picture is generated by background.py: nothing was downloaded and no cover art was invented. ### Speech-aware silence removal keeps the breaths ![Speech-aware silence removal keeps the breaths](demos/silence_speech_aware.gif) ```bash python3 scripts/silence.py demos/out/silence_speech_aware_breathy.mp4 --threshold -35 --min-silence 0.2 --margin 0.1 --list --json python3 scripts/silence.py demos/out/silence_speech_aware_breathy.mp4 --threshold -35 --min-silence 0.2 --margin 0.1 --preset veryfast -o demos/out/silence_speech_aware_naive.mp4 python3 scripts/silence.py demos/out/silence_speech_aware_breathy.mp4 --threshold -35 --min-silence 0.6 --speech-aware --list --json python3 scripts/silence.py demos/out/silence_speech_aware_breathy.mp4 --threshold -35 --min-silence 0.6 --speech-aware --preset veryfast -o demos/out/silence_speech_aware_after.mp4 ``` **Look for:** A 0.25 s breath sits inside a sentence and a 1 s pause sits between sentences. Left uses a --min-silence low enough to catch the breath too, so plain detection cuts both; right is --speech-aware at the ordinary --min-silence -- its finer 0.12 s floor still measures the breath, but the breath is shorter than --min-silence so it is kept, and only the real sentence-boundary pause is cut. ### Korean filler words cut by the built-in list ![Korean filler words cut by the built-in list](demos/filler_korean.gif) ```bash python3 scripts/silence.py demos/out/filler_korean_speech.mp4 --filler --words demos/out/filler_korean_words.json --filler-list --json python3 scripts/silence.py demos/out/filler_korean_speech.mp4 --filler --words demos/out/filler_korean_words.json --margin 0 --preset veryfast -o demos/out/filler_korean_after.mp4 ``` **Look for:** Five timed words, two of them the hesitations 음 and 어 (the low hum). Right is silence.py --filler with the transcript's language read as Korean: both hesitations are gone, and 음악은 (music), which starts with the same syllable, stays -- a filler matches a whole word, never a prefix. ## Delivery & checks ### Reels export, then checked ![Reels export, then checked](demos/export_reels.gif) ```bash python3 scripts/export.py demos/out/fixtures/motion.mp4 --preset reels --fit crop -o demos/out/export_reels_after.mp4 python3 scripts/check.py demos/out/export_reels_after.mp4 --platform reels --json ``` **Look for:** One command produces the 1080x1920 deliverable; check.py then reports the spec row by row and exits non-zero on a FAIL. ### TikTok delivery template ![TikTok delivery template](demos/template_tiktok.gif) ```bash python3 scripts/render.py demos/out/fixtures/motion.mp4 --template tiktok --cues demos/out/fixtures/cues_en.txt --fast -o demos/out/template_tiktok_after.mp4 ``` **Look for:** One command turns the master into the 1080x1920 deliverable: reframe, burned-in captions kept clear of TikTok's own UI, loudness to -14 LUFS, the tiktok export preset and check.py's platform rows. ### Proxy vs. master, shown at the same size ![Proxy vs. master, shown at the same size](demos/proxy.gif) ```bash python3 scripts/proxy.py demos/out/fixtures/mandel.mp4 --width 320 --quality 34 --fps 12 -o demos/out/proxy_after.mp4 ``` **Look for:** Both halves are scaled to the same cell, which is the honest comparison: the proxy is smaller on disk and cheaper to decode, not smaller on screen. The softness is the point of it. ## Projects & inspection ### Whole edit from one project file ![Whole edit from one project file](demos/render_project.gif) ```bash python3 scripts/render.py demos/out/fixtures/project.json --fast python3 scripts/render.py demos/out/fixtures/project.json --export-timeline demos/out/render_project_timeline.otio ``` **Look for:** Clips, a transition, captions and a title card described as JSON and rendered in one pass. ### A keyed video layer in a project ![A keyed video layer in a project](demos/render_pip.gif) ```bash python3 scripts/render.py demos/out/fixtures/project_pip.json --fast ``` **Look for:** The 16:9 master is reframed to 9:16 on a blurred background, then a green-screen clip is keyed and composited at the bottom. The overlays stage runs after fit, so "scale": 540 is the full width of the delivered frame. ### Contact sheet ![Contact sheet](demos/contact_sheet.gif) ```bash python3 scripts/look.py demos/out/fixtures/mandel.mp4 --tiles 4x3 --width 960 -o demos/out/contact_sheet_sheet.png ``` **Look for:** Twelve timecoded frames in one PNG: the fastest way to confirm an edit landed where it should. ### Scene detection into a highlight reel ![Scene detection into a highlight reel](demos/scenes_highlights.gif) ```bash python3 scripts/scenes.py demos/out/fixtures/shots.mp4 --highlights 2 --min-scene 1 --edl demos/out/scenes_highlights_picks.txt python3 scripts/cut.py demos/out/fixtures/shots.mp4 --segments 0.00-2.00,2.00-4.00 --accurate --preset veryfast -o demos/out/scenes_highlights_after.mp4 ``` **Look for:** scdet finds the hard cuts, scenes.py ranks the scenes and writes the ranges as an EDL, and cut.py --segments is what turns that proposal into a reel. The ranking is a proxy for interest, not a judgement of it. ### Chapter marks written into the container ![Chapter marks written into the container](demos/metadata_chapters.gif) ```bash python3 scripts/metadata.py demos/out/fixtures/mandel.mp4 --chapters demos/out/fixtures/chapters.txt --title 'Demo episode' -o demos/out/metadata_chapters_tagged.mp4 ``` **Look for:** Nothing in the picture changes -- every stream is copied bit for bit. The right half is the chapter list read back out of the file with ffprobe and drawn as the timeline strip a player would show. ### Chapters proposed from measured structure ![Chapters proposed from measured structure](demos/metadata_auto_chapters.gif) ```bash python3 scripts/metadata.py demos/out/fixtures/shots.mp4 --auto-chapters --min-chapter 1 --from scenes -o demos/out/metadata_auto_chapters_auto.mp4 ``` **Look for:** The left strip is the file with no markers; the right is the same file after --auto-chapters, whose timestamps come from the measured pauses and scene cuts and whose titles are deliberately 'Chapter 1..N' -- the skill proposes where, the caller says what. ### Shots labelled static / pan / motion ![Shots labelled static / pan / motion](demos/scenes_shots.gif) ```bash python3 scripts/scenes.py demos/out/fixtures/shots.mp4 --shots --json ``` **Look for:** scenes.py --shots measures an optical-flow proxy per detected shot and reports a label for each -- a measurement, not an edit. The strip shows the three shots in the fixture with the labels actually returned. ## Inspection These tools answer a question instead of changing a picture, so there is nothing to put a before and an after next to. Each one prints a table, a JSON document or writes an HTML file; the command is the demo. ### probe.py ```bash python3 scripts/probe.py demos/out/render_project_after.mp4 --compact ``` **No picture:** duration, fps and VFR suspicion, frame size, codecs, pixel format and colour tags, audio channels and sample rate -- one JSON document, or one line per file with --compact. ### verify.py ```bash python3 scripts/verify.py ~/Footage --quick --report verify.md ``` **No picture:** runs the whole toolchain over your own files and prints a PASS/FAIL table, one row per tool per file. Synthetic fixtures cannot show what a real phone container does; this can. ### batch.py ```bash python3 scripts/batch.py ~/Footage --recipe batch.json --dry-run ``` **No picture:** applies one recipe to a folder with a content-hash cache, printing what it would run, what it skipped as unchanged, and where each output lands. ### check.py --content ```bash python3 scripts/check.py final.mp4 --platform youtube --content ``` **No picture:** adds content rows to the platform check: black frames, frozen frames, the share of silence, and short dropouts (digital silence for 20 ms to 1 s with sound on both sides, the mark of lost audio at a join), single-frame flashes (one frame unlike both of its neighbours) and a first frame of one flat colour. The dropouts and flashes rows list the first five times; a row that does not pass names a fix. ### check.py --levels ```bash python3 scripts/check.py spot.mxf --platform broadcast --levels ebu-r103 ``` **No picture:** counts the luma and chroma samples outside the legal range in every frame (bt709 16-235 / 16-240, or EBU R 103's -1..103 % luma) and adds a legal levels row with the worst frame's time and min/max Y/U/V; the JSON has the shares per window. color.py --legalize clamps after the 4:2:0 conversion and measures the file it wrote the same way -- a change no preview would show, which is why it is a number here and not a GIF. ### check.py --expect-changed ```bash python3 scripts/check.py new.mp4 --platform custom --compare old.mp4 --expect-changed 12-18 ``` **No picture:** after a small re-render, says whether anything changed outside the span the edit was meant to touch: both pictures are subtracted frame by frame on a 32x18 grid, and a change_scope row FAILs with the time and place of a change nobody declared, WARNs when the declared span did not change, and FAILs when the two files could not be compared. A 200x60 box on a 1080p frame leaves the whole-frame SSIM above 0.99; this row finds it. ### report.py ```bash python3 scripts/report.py --before raw.mov --after final.mp4 -o report.html ``` **No picture:** a single self-contained HTML page: before/after contact sheets, loudness, the platform check table and the exact commands. It is a file to open, not a picture to embed here. --- Missing a feature you use? A `feat` PR is expected to add a demo here and in `demos/build.py` -- see [CONTRIBUTING.md](../CONTRIBUTING.md).