Skip to content

Proofs roadmap

Each major feature shown working on real footage that does not already show the result, before and after, on the site’s features page (the Before and after section, apps/site/src/components/Proofs.astro, fed by apps/site/src/lib/proofs.ts). A proof is what a visitor can check with their own eyes; a sentence on a page is a claim.

Status: step 0 done (the stage, the manifests, the script, the landing’s strip). Shown on the site: a talk edited by the assistant, three shorts from a long video, colour grading, keying, a photo in 2.5D, a title behind a person, a background removed, matching two shots, skin retouch, tracking, a vertical reframe, slow motion, upscaling, a clean voice, word-by-word captions, silences cut. One draft, with what it waits for: stabilisation (a handheld shot of a still scene; the clips tried are steady). Most proofs found and fixed a defect of the feature they show (see each box).

The first grading test was run on a clip that was already clean: the difference did not show, and the user did not believe it. The same test on flat log footage showed it at once, and it also found a real defect (out-of-gamut colours turning cyan) that a clean clip would never have shown. So a proof does two jobs: it convinces, and it tests the feature on the material it exists for.

  • A source that has the problem. Flat log for grading, a shaking hand for stabilisation, a noisy room for voice cleanup, a horizontal shot for a vertical reframe. Never footage that already looks done.
  • Untouched, and free to show. Downloaded as published, its licence read on the day and recorded with the author and the page. Accepted when the licence allows modification and website use: Pexels, Pixabay, Unsplash (photos), CC0, CC BY and CC BY-SA (the derived clip is then published under CC BY-SA, said under it), public domain (NASA, the Prelinger Archives on the Internet Archive, LibriVox for voices), the Blender open films (CC BY). Each licence is re-read before use: they change.
  • Everything in Tramme. No step outside the app: no hand retouch, no other grader, no re-encode with filters. The settings used are saved with the proof and shown under it, so anyone can redo it.
  • The same frame, the same time. Before and after compared at the same instant, at the same size, encoded the same way.
  • Representative, not picked. The proof uses the default or a stated setting; the clip is chosen for its problem, not for where the tool happens to shine. A feature that fails its proof is fixed first, as the grading was.
  • A measure where there is one. Stabilisation: the residual motion in pixels per frame. Voice cleanup: the noise floor in dB. Silences: the seconds taken out. Upscaling: a crop at 200 %. Written beside the proof.
  • People shown with respect. Pexels and Pixabay forbid showing identifiable people in a bad light: a retouch proof shows a subtle change on a neutral portrait and says it is retouched. No proof alters what a real person says (no lip sync, no dubbing of identifiable people).
  • Light. Per proof: stills at most 800 KB each, a clip at most 4 MB (WebM VP9 first, MP4 fallback, a WebP poster), no sound unless the proof is about sound. Nothing of a proof loads before the visitor picks it.

Today each proof is laid out in full, one under the other: a slider beside its text, then the clip across the page, about one and a half screens each. That holds two or three proofs; at fifteen it is a twenty-screen scroll.

The section becomes one stage of fixed height, whatever the number of proofs:

  • a row of the features’ names above it (scrolling sideways on a phone), the chosen one shown, the others loaded only when picked;
  • in the stage, one frame with a switch: Picture (the slider) or Video (the clip), instead of a second block below;
  • beside it (under it on a phone), the text, where the tool is, the settings, the measure and the credit;
  • a link per proof (/features/#proof-<id>), so a page or a post can point at one;
  • three kinds of comparison: a picture (the slider), a video (the clip with its line sweeping, or two clips side by side in sync, for a reframe or a slow motion), a sound (one player, a Before / After switch at the same position, its waveform drawn).

The landing page gets a strip of three proofs, each a poster that opens its proof on the features page.

  • A proof’s manifest, proofs/<id>.json: the source’s page, its file’s address, its licence and its page, its author, its SHA-256, the part used (in and out), the Tramme document used for the after (with its settings), the measure, and the texts.
  • node scripts/proofs.ts <id>: downloads the source into out/proofs/<id>/ (never committed; the hash checked), renders before and after with the command line (tramme still, tramme render), measures, and writes the stills to apps/site/src/assets/proofs/ and the clip to apps/site/src/videos/ at the site’s sizes. Run again, it gives the same files.
  • lib/proofs.ts reads the manifests instead of holding the texts (the texts in both languages stay in the manifest).
  • The stage described above, with the picture, video and sound kinds; checked on a phone and a computer, light and dark, keyboard alone.
  • The grading proof moved to the manifest and the stage, its files unchanged.
  • The strip of three proofs on the landing page, nothing loaded before it comes near (the landing’s performance rule).
Proof Tool, where in the app Source with the problem Where to look Shown as Measure
Keying Chroma key (fx.key: Effects, +, Effects), its matte view A green-screen plate with hair and motion The Blender film Tears of Steel, its green-screen plates (CC BY); Pexels “green screen” Picture: the plate, then keyed over a background; the matte —
Cut-out without a green screen the cutout tool, text behind the person A person walking in front of a busy background Pexels, Pixabay Video —
Background removed from photos remove-bg Product photos on a cluttered table Unsplash, Pexels Picture over a checkerboard —
A photo in 2.5D the depth tool, fx.parallax A landscape or a street photo with clear depth Unsplash, Pexels Video: the still photo, then the camera moving in it —
Matching two shots Match the look of the first clip (clip menu) Two shots of one scene from two cameras, different white balance Pexels (same author, same place), the Tears of Steel plates Picture: the two shots side by side, before and after The mean colour difference (ΔE) between them
Skin retouch the skin tool, fx.smooth A neutral close-up portrait, natural skin Pexels, Pixabay (a portrait, shown with respect) Picture, a crop at 100 %, labelled retouched —
  • Keying (a Pexels green-screen take; it led to the colour-difference matte)
  • Cut-out without a green screen (a Pexels walk along a river; it led to footage holding its frame across the motion blur’s shutter)
  • Background removed from photos (a Pexels camera on wooden planks: the camera kept, the plant left out, the wood’s colour taken out of the edge)
  • A photo in 2.5D (a Pexels lake photo; it led to the parallax following the camera’s pan, mirrored edges, the framing for the drift and no depth of field by default: the blur read as a softer picture, not as depth)
  • Matching two shots (two Pexels shots of the Golden Gate Bridge by two people; it led to the matching keeping black near black, a camera’s cast matched whole and the brightness and contrast only so far, and to the tool waiting for the preview to load a clip just added; ΔE of the mean colours 10.4 to 8.6, the rest is what each shows)
  • Skin retouch (a Pexels neutral portrait in 4K, shown at 100 %, the default strength 0.3, labelled retouched)
Proof Tool, where in the app Source with the problem Where to look Shown as Measure
Stabilisation the stabilize tool (the video’s Stabilisation) A handheld walk, clearly shaking Pexels, Pixabay (“handheld”, “walking POV”) Video, side by side in sync Residual motion, px per frame, before and after
Tracking track, then callout A car, a bike or a runner crossing the frame Pexels street footage Video: a label that follows Drift of the label, px
Vertical reframe the shorts reframe, focus following the subject A horizontal shot whose subject moves sideways Pexels, Pixabay Video: the 16:9 source beside its 9:16 —
Slow motion the long jobs’ frame interpolation A fast action at 25 or 30 fps (a jump, water) Pexels, Pixabay Video, side by side: frames repeated against frames made —
Upscaling the long jobs’ upscaling (the engine in use is named) Low-resolution archive film The Prelinger Archives (public domain) Picture, a crop at 200 % —
  • Stabilisation. First try (a walk filmed looking down, Pexels 4405576): no gain, the jitter 28.1 px before and 46.8 after a 1.5x zoom (node scripts/proof-shake.ts); the legs crossing the frame move with the ground the tracker follows. Needs a handheld source of a still scene, and the tool to set moving subjects aside (outliers by region). The manifest stays as a draft. A second search (2026-11-13): four church interiors and a clip sold as handheld measured 0.1 to 1.5 px, steady; the free libraries describe no camera movement, so each candidate has to be downloaded and measured.
  • Tracking (a Pexels red coupé crossing a street, filmed by hand; it led to the first frame of a render waiting for a 4K source to open, and to the clips held under 4 MB by a two-pass encode)
  • Vertical reframe (the Pexels walk along a river, the man drifting across the 16:9, the 9:16 frame following him)
  • Slow motion (a Pexels border collie leaping for a frisbee in 4K at 30 fps, four times as many frames made by RIFE in the browser; it led to the frames being made at the film’s size, in overlapping tiles above 1080p, after a whole 4K frame hung the graphics card, to a fallback on the processor when the card stops, and to the frames RIFE makes losing a rounding that brightened them by one level)
  • Upscaling (a 640 × 480 photo from Wikimedia Commons, public domain, enlarged by Real-ESRGAN in the browser; a video stays with Lanczos, said under it)
Proof Tool, where in the app Source with the problem Where to look Shown as Measure
Captions, word by word Transcribe, then captions A person speaking, clear sound NASA (public domain), Wikimedia Commons (CC BY, CC BY-SA) Video with sound, captions in time Words timed within 100 ms, on a sample
Silences and hesitations cut silences, fillers A talk with pauses, “um”s and restarts NASA briefings, Commons interviews Video with sound, the raw take beside the cut one Seconds removed, cuts made
A clean voice denoise A voice recorded in a noisy place (street, wind, hum) Freesound (CC0, CC BY), Commons, LibriVox (some readings are noisy) Sound: the Before / After player Noise floor, dB
  • Captions, word by word (a NASA interview, public domain: transcribed by Whisper in the browser, the spoken word lit)
  • Silences and hesitations cut (a NASA HERA Groundlink, two crew members answering live, the voice recorded alone: 60 s to 55.4 s, 7 cuts). The first try, a NASA interview with music under the voice, found three defects of the tool, fixed (only 0.24 s was taken from each pause; a clip showing part of its file got an invalid document; word timings stretched over the pauses hid them, the sound is now heard by the voice detector).
  • A clean voice (a street interview from Wikimedia Commons, CC BY 3.0: background noise from −41 to −57 dB, the voice’s level unchanged)

These run the assistant, so they cost the user’s quota: each waits for the user’s go, and its cost is written with it.

Proof What runs Source Shown as
A talking video, edited The assistant on a raw take: cuts, captions, inserts that show what is said, a grade A public-domain or CC talk (NASA, Commons), 1 to 2 minutes Video: the raw take beside the finished film, the same 30 s
Shorts from a long video shorts on a talk A CC talk of 10 minutes or more Three vertical shorts beside the long video’s timeline
  • A talking video, edited (a NASA raw take, 60 s to 48 s, 12.7 min and $1.84; the first runs lost their last edits to a save conflict, fixed: saves go one after the other)
  • Shorts from a long video (a NASA talk of 5 min to three shorts of 42, 21 and 33 s, 42 minutes and $1.04; it found two defects, fixed: a tool call longer than 15 minutes through the companion was dropped, the editor now says it is still at work; the command line’s text checks measured heavy capitals too narrow, a keyword running to the edges passed)
  • Lip sync and dubbing: they change what a real person says; no proof on someone who has not agreed to it.
  • Voice and music apart: blocked (docs/editing-roadmap.md: the one MIT export of HTDemucs measured does not put speech in its vocals).
  • HDR export: waits for a canvas above 8 bits.
  • Features without a before and after (transitions, packs, the studio’s notes): shown by the examples and the showreel, not here.
  1. Sources are never committed: the manifest holds their address and hash, out/proofs/ holds the files.
  2. A proof that would need a step outside Tramme is not published; the missing step becomes work on the feature.
  3. The texts say what was done and what was not, in plain words, and name the engine when a better one is planned (interpolation, upscaling).