Chapter 30 · Finishing, captions and graphics
Part 6 · Edit, finish and delivery
Finishing is where a generated film can be improved or quietly damaged. An upscaler can invent a new face, a grade can be laid on the wrong reading of a file, a caption timed to the wrong audio drifts off the lips, and a title generated as pixels can never be retyped. This chapter sets the order of finishing work and the standard for each step: the upscale decision, colour and texture, sound and loudness, Arabic captions and Arabic in the frame, the versions a delivery needs, and graphics built natively or generated as pixels.
In this chapter
The finishing order: lock, conform, repair, upscale only if needed, colour, texture last
Probing and conforming files that disagree with their labels
Deciding whether to upscale, by two labelled previews in motion
Colour: interpret the source; never lay a camera-log transform on generated media
Grain and halation: added in post, restrained, last, never prompted
Sound in the finish, and loudness measured against the named specification
Captions from approved copy, timed to the audio that shipped; Arabic in a frame, always exact
Aspect ratios and versions
Motion graphics: native, or generated as pixels
Before you start. Chapter 29 makes and locks the cut; apart from probing and conforming files, which begin the moment a file arrives, nothing here starts before the lock. Chapter 10 teaches the two layers of text: in-world text may be generated, editorial text never is. Chapter 14 sets formats and frame rates, Chapter 21 says who owns each sound, and Chapter 22 the routes to a speaking face, which decide how its captions are timed. Chapter 32 holds the Premiere Pro and upscaler settings, and Chapter 31 hands over what this chapter finishes.
30.1 The finishing order
Finishing runs from a locked picture to a set of versions another editor can open. The order is the point: each step is judged before the next, and nothing that costs money happens before the cut is accepted. No maker documents editorial finishing; the order below is professional practice, and its last two picture steps follow the Mastorna trailer's own finishing note (30.5).
Word | In this book |
Finishing | everything done to picture and sound after the cut is locked: conform, repair, upscale, colour, texture, text, sound and versions |
Conform | converting a file deliberately to the sequence's frame rate or frame size |
Native original | the file as generated and accepted, kept untouched beside any transformed copy |
Upscale | making a larger file from a smaller one; the upscaler invents the new detail |
Input interpretation | how a file's colour is read: its transfer curve, primaries and range, taken from its tags and confirmed by eye |
LUT | a look-up table: a preset colour transform |
Texture | film grain and halation, added in post |
Halation | the soft glow around bright highlights that a film base scatters |
Canonical line | the approved text of a line of dialogue, the authority for every caption |
MOGRT | a Premiere motion graphics template: an editable graphic whose text can be retyped |
Figure 30.1 — The finishing order. Step 1 comes first: nothing after it is spent until the picture is locked, and upscaling is decided only on the intervals the cut uses. Texture comes after colour, and last among the picture steps.
# | Step | Check before the next step |
1 | Lock the picture | every item linked to its source and interval; every speed change, reframe, stabilisation and interpolation listed; unresolved items stop the lock (29.14) |
2 | Probe and conform | size, frame rate, frame count, duration and audio stream against the request; rate and raster conformed on purpose |
3 | Repair named defects | in used intervals only: a trim, a local fix, an audio bridge or a child shot before any global transform |
4 | Decide on upscaling | two labelled previews of one hard moving interval, judged in motion; the native original kept |
5 | Colour | the source interpreted, shots matched, skin and product protected, the look applied separately, viewed through each output transform |
6 | Texture | grain and halation added last, at final resolution, restrained and matched across every element |
7 | Text and graphics | from approved copy and real assets, on their own tracks; Arabic checked in the real player |
8 | Sound | dialogue, room tone, effects and music balanced; loudness, true peak and downmix measured on the final export against the named specification |
9 | Versions | every target overlaid upstream; each shot classified; each version with its own quality check |
10 | Final exports and handover | probed, reopened, compared; the package (Chapter 31) |
The finished grade goes through every step, in this order. The test grade goes through almost none. When the aim is only to prove a route cheaply, finishing shrinks to a hard-cut assembly slated as a draft (29.6), with no grade, upscale or graphics, and one reference render. It is labelled a test and never presented as the model of a finished version.
Lock first, spend later, and let the texture be the last thing the picture receives. Every credit spent finishing a shot the cut later drops is lost, and a grade or an upscale laid on a moving cut has to be laid again.
Work only on the locked cut's intervals, the ones the cut uses (29.14), not the whole files.
Repair before you transform. Repair by trim, shorter interval, audio bridge or child shot (29.11) before any global transform. A global transform is the wrong tool for a defect a shorter interval would have removed.
Log every transform, and remember that finishing proves nothing about a generator. A shot that only post could rescue is a shot the route could not make.
A finishing decision is the director's. The assistant prepares the comparisons and records the settings; the director chooses.
30.2 Probe and conform
Probe every file before it enters the project, and keep the original download untouched. To probe is to measure a file with a tool such as ffprobe instead of trusting its label, its file name or the job record. Files disagree with their labels: a request for 720p has come back at 1284 × 716, a request for 1080p at 1912 × 1080, and one clip has come back at 30 frames a second among clips at 24. Record the provider's label separately from the measured file, and leave native or upscaled provenance marked unknown until it is known.
Probing a returned file
Let the download finish, then keep that file untouched as the original; anything you change is a copy.
Probe it and take its checksum. Record duration, frame size, the exact frame rate (a rational rate such as 24000/1001 as it is, not rounded to 23.976), the frame count, the audio streams and the checksum.
Compare the record with the request: size, rate, length, and whether an audio stream is present that the settings switched off.
Label any difference on the file's record ("audio stream present", "1284 × 716") and move on. A prompt cannot fix what the platform did to the file.
Conform rate and raster deliberately, and never stretch a raster to hide a mismatch. A plate a few pixels off the sequence size is scaled or cropped on purpose: measure what each costs before you decide.
Determine a file's real cadence from the file, and never assume a resolution or a mode implies a frame rate. Avoid duplicated, dropped or interpolated frames unless they are approved and disclosed, and verify duration, motion and audio sync after the encode.
Convert a delivery codec to an editing codec first when the editing application or the lip-sync route cannot read it; a 10-bit HEVC file is the usual case. The tool sections say what each model returns (Chapters 16 to 18).
Plan handles from the probed file (29.11): the seconds you asked for, the frames you got and the interval you select are three different numbers, and nothing exists beyond a file's ends.
30.3 Repair, then decide whether to upscale
Finishing starts after the picture is locked, on its used intervals. Repair named defects first, by the cheapest lever that works (29.12). Then decide, once, whether the delivery needs more pixels than the files have.
Treat upscaling as a labelled decision, never a default. There is no rule of the form "upscale all 1080p to 4K". The delivery size decides whether a clip is upscaled; a comparison decides how. Keep the native original, and label the copy "upscaled (tool, version, date)".
Price the alternative first: where a model can generate at the delivery size, compare that price with an upscale's, read on the day (34.2).
Upscale from neutral footage. If the delivery needs it, upscale the used intervals before colour, so that the grade stays a live decision in Premiere and a change of grade never means running the upscaler again. Grade at the delivery resolution afterwards. This is a default: no maker states an order, and a comparison on the interval in question outranks it.
Figure 30.2 — An upscale is judged as a comparison, never on its own. Pixel dimensions are not a verdict.
Judging an upscale
Pick one hard moving interval from the locked cut: skin, hair, edges, the label, Arabic letters.
Make two labelled previews with the same grade and framing: the native size, and the upscale at the delivery size.
Judge each transformation on its own (denoise, sharpen, upscale), never combined: combining them hides the cause of waxy skin, edge shimmer or lost grain.
Compare in motion at the viewing size, and the label and any Arabic letter by letter.
Treat any change of identity or lettering as a failure, and prefer a stable, slightly softer result to a nominally higher-resolution one that damages the product.
Keep the native original, label the copy, and record the tool, version and settings. The director chooses.
A still proves nothing about shimmer, and the delivery size decides, not the pixel count: compare in motion, on the hardest interval, never on a flattering frame.
Numeric settings from a third party's guide are not rules. A "recover detail" range or a grain strength belongs to one tool's version and one guide's footage. The comparison is the method; the settings that Premiere's panel and the platform's upscalers expose are in Chapter 32.
A frame-rate boost is an upscale of time. Invented frames are drawn between real ones, and lips, hands and edges show it first. Disclose every boost, inspect it, and check the mouth on a speaking shot frame by frame.
For a lip-synced shot, upscale after the sync, or make the plate at the delivery resolution in the first place. Lip sync keeps the plate's size (Chapter 24); which of the two is better for a given delivery is not yet known, so test both on the hard interval.
Research exemplar · beauty and personal care · a brief to the assistant editor: two labelled previews of one interval · copied exactly from the September 2026 creator research; instructional, never run
Example 30.1 — comparing a finishing candidate
Prepare two labelled finishing previews from the same selected interval and color interpretation: the native-resolution version with a restrained grade, and the candidate upscale with matching framing and grade. Compare skin texture, moving hair, glass edges, the label, shadow noise and fine highlights throughout the turn. Record the tool settings and any identity or lettering change. Recommend a version for the actual delivery size; do not use pixel dimensions alone as the quality verdict.The brief fixes grade and framing, so the only variable is the upscale and any change is the upscaler's.
It names what to look at (skin, moving hair, glass edges, the label, shadow noise, fine highlights) and asks for them through the turn, not on a frame.
Its last sentence keeps size out of the verdict: recommend for the delivery, never by pixel dimensions.
If the label shimmers, deliver native and state the limit; where the delivery really needs the size, a separate origination at the larger size is the second alternative.
The chosen version sits, labelled, beside the native original, and the tool, provider and settings go into the run record (8.7) and the finishing log.
30.4 Colour
Colour comes after the upscale decision and before the texture, and one rule governs it: interpret the source before you touch the look. Generated files carry colour metadata that may be incomplete or misleading, so the first step is a probe of the file's tags and a look at the picture, not a preset.
Grading generated footage, in order
Probe the file's colour tags, and establish a trusted way of viewing the picture.
Choose and document the input interpretation for each source, and record the uncertainty where the tags and the picture disagree.
Normalise the shots into the working space.
Match across the scene: identities, product and any composited elements, against the approved masters.
Apply the creative look as a separate step, after the matching.
View through each output transform the delivery will use, and check scopes and images.
Render, and record the interpretation, working space, look, output transform, range, bit depth, codec, version and checksum.
Never give generated media a camera-log transform because a LUT looks good. A LUT made for camera-log footage expects the flat, low-contrast image a log camera records. A generated clip was never recorded that way, so a log transform distorts it, however good one frame looks.
A colour-management framework is not a look. ACES, for instance, says how each source is read and how the result is shown; the look is a separate creative decision, and the two are recorded separately.
A digital look inherits nothing from a physical stock. A LUT or emulation is specific to its vendor, version and pipeline, and carries no universal warm or cool character.
Protect skin and product by matching against their masters, not by a fixed rule. There is no universal skin hue band, no beauty smoothing by default and no regional palette. Evaluate each approved identity under each shot's motivated light and output transform, correct only where needed, and keep natural variation and texture.
Match colour across shots from different routes without forcing the backgrounds to agree. The product's colours must agree; a background that differs because its light differs is not an error. An amber bottle is matched to the approved product master (hue and density, cap, label, stone neutral), with the glass highlights protected; when one shot drifts warm, return to the master and correct that shot alone.
30.5 Texture: grain and halation, last
Add grain and halation in post, restrained, at the final resolution, after the colour, and never ask a generator to "fake old film". Texture is applied only if the look approves it, and after grading, not before. The Mastorna trailer's finishing note states the method in five lines:
Final artifact should look like a pristine restored black-and-white 35mm trailer print.
Add fine grain and restrained halation in post rather than asking models to “fake old film.”
Keep registration stable; only minimal gate movement if any.
Do not use heavy scratches, mold, cigarette burns, or nostalgia-damage clichés.
Preserve photochemical softness and silver-range blacks while avoiding digital HDR crispness.
The division of labour is the lesson. The look is made in post; the generator is not asked to imitate film. The fourth line forbids the damage clichés by name. Black-and-white, a restored print and 1966 belong to that trailer; the rest transfers.
A still's finish block and the film's delivery texture are two layers, and each has one owner. The still's finish block (9.5) describes capture: the stock's character and tonal response, where highlights bloom, what the lens does at the edges, with grain fine and restrained as a property of the photograph. It never asks the model to age or damage the image. The delivery texture, the grain and halation the audience sees, is a post decision: made once, matched across every shot and graphic, added last at the final resolution, and never prompted as an effect.
Texture settings are specific to the application and the shot. Choose grain size, channel behaviour, intensity and temporal character by the final resolution and the approved look, match every composited element, add halation only where the look calls for highlight glow, and record tool, version and settings. There is no universal percentage.
Do not count on grain to repair a cut. It does not fix a shot whose identity, light or motion disagrees with its neighbours. Fix the shot. If the texture looks pasted on, match its size and character to the final resolution and to every composited element; if halation glows where the intent does not call for it, remove it from those highlights.
30.6 Sound in the finish, and loudness
The sound is built in the edit (Chapters 21 and 25); in the finish it is sorted, balanced and measured.
Strip the native tracks you did not choose. Generated clips can carry a soundtrack nobody asked for: a stereo bed under a silent request is a known case. Mute that layer and rebuild dialogue, room tone, ambience, Foley and music from approved isolated assets, using restoration only where it does not damage speech.
Carry the cut's sound rules into the finish unchanged. The audio that ships owns the timing: the locked take on the exact-take route, the clip's own audio on the new-performance route (Chapter 22). One continuous room bed runs under every cut of a scene, and the track lanes of 29.2 hold.
Name the delivery specification before you measure (8.2 asks for it in writing at the start), and never type a loudness figure into a prompt or a render setting. ITU-R BS.1770 defines how loudness is measured; it is not a target. EBU R 128 and ATSC A/85 belong to named broadcast scopes, and cinemas, streaming services, social platforms, web embeds and client reviews may each require different targets, peaks, channels and codecs. Measure the final export, and its downmix, against the specification the delivery owner names, keeping true peak and dynamics in view. Measure the finished mix as a whole, without emphasis on speech, music or effects.
The published specifications (checked 30 Sep 2026). Two families cover short commercial work.
Specification | Scope | Loudness target | Peak | Notes |
EBU R 128, version 5 (Nov 2023) | broadcast programmes; the definition includes advertisements, trailers, promos and interstitials | −23.0 LUFS integrated; ±0.2 LU where quality control checks it, ±1.0 LU for live or impractical cases | not above −1 dBTP | measured over the whole programme |
EBU R 128 s1, short-form (Aug 2020) | short-form content | −23.0 LUFS (±0.2 LU) | not above −1 dBTP | short-term loudness not above −18.0 LUFS; loudness range "not applicable" |
EBU R 128 s2, streaming (Nov 2023) | programmes distributed by streaming | recommends streaming unchanged, at −23.0 LUFS | permits a distribution level above −23.0 LUFS, with an interim range of −20.0 to −16.0 LUFS, only where the broadcaster controls the dynamic treatment, and says to switch to −23.0 once devices give enough headroom | |
ATSC A/85, edition 2026-07 (a Recommended Practice), Annex M | United States, short-form content | −24 LKFS for content delivered without metadata, with a tolerance of up to about ±2 dB | not above −2 dBTP (about ±0.5 dB of meter tolerance is accepted) | short-form not louder than the long-form it accompanies; for streaming delivery only, a single target between −23 and −27 LKFS is recommended |
If the client names no specification, use a working default and ask. For a European-style broadcast spot: −23.0 LUFS integrated, short-term not above −18.0 LUFS, true peak not above −1 dBTP. For a US-style spot: −24 LKFS, true peak at or below −2 dBTP. For online-only versions: the platform's or the client's specification, and where none exists, deliver the broadcast version and ask before making a louder one.
Egyptian, Saudi and Gulf broadcasters, cinema advertising and the social platforms: no publisher's figure was found (checked 30 Sep 2026), and none is given here until the owner's own page has been read on the day. EBU supplement 4 is not a cinema-advertising figure: it guides adapting dynamic cinematic content for broadcast and streaming. Ask the delivery owner for the specification in writing, and record the answer with its date.
A voice that "peaks around −3 to −6 dB" is not a specification: broadcast specifications limit true peak and set programme loudness. The balance of voice, music and effects is a judgement about intelligibility and the dynamic arc, and normalisation is no cure for bad editing or noise.
Record one specification per deliverable, measure each on its own, and keep the measurement. Social, broadcast and cinema or review versions each have an owning loudness, peak, channel and codec contract. A version that measures fine and sounds wrong in mono needs its phase and channel routing checked and its downmix measured again. Record the method, target, result, true peak, channel layout, sample rate and codec beside every final export, with a hash (Chapter 31); a review copy normalised for listening measures nothing about a delivery.
30.7 Captions, and Arabic text
A caption is editorial text: it sits on the image, and it is built in the edit from real text, never generated. Titles, subtitles, supers, end-cards, watermarks and logo lockups are composited in post from real assets, and the closing exclusion of a generated frame, "no overlaid text, subtitles, logos, or interface graphics", bans exactly that layer (10.12). Three things decide whether an Arabic caption is right: which text it uses, which audio it is timed to, and how the text engine sets it.
The text: the canonical line
Start captions from the approved copy, never from a transcript and never from the voice engine's respelling. Record the spoken dialect and the subtitle language as separate fields, so that a Cairene line is spoken in Cairene and subtitled in whichever language the deliverable needs. Do not convert colloquial speech to Modern Standard Arabic unless the localisation decision says so.
Do not plan transcript-led cutting of Arabic dialogue in Premiere. Its own Speech to Text does not list Arabic (32.2), so an Arabic transcript comes from the approved script or from another tool, and every timing it gives is checked against the waveform.
The timing: the audio that shipped
Time every caption to the audio that shipped. On the exact-take route the locked take ships and the captions follow it. On the new-performance route a video model performs the line, re-times it, and its own audio ships; captions timed to the take it was given would run early against the lips. A line that took 3.9 seconds in the take may arrive with its first sentence ending at 2.3 seconds and a pause before the second, and the caption break belongs at that pause. After a late change of line, the new take is the authority. Captions are re-issued whenever a line or a take changes.
Figure 30.3 — The timing authority for a caption. The words always come from the approved copy; only the timing changes with the audio that shipped (Chapter 22 describes the routes).
Building captions
Take the approved copy, never the engine's respelling.
Time it to the audio that shipped (Figure 30.3).
Set Arabic direction, shaping and order, and test them in the real player.
Re-issue the captions whenever a line or a take changes, and deliver a caption file for each version with the approved text beside it.
Template: Arabic captions (written for this book; not run)
Build Arabic captions for [shot] from the canonical line [line], not the engine spelling. Time them to [the shipped audio: the locked take / the clip's own audio]. Break at phrase boundaries; keep dialect as spoken unless the localisation decision says otherwise. Return the caption file and list the timings you measured.The setting: Arabic in Premiere
Build every title, caption and end card in Premiere from the approved copy. Pre-render cards with an alpha channel in the typeface the project keeps for retyping. Keep captions on their own caption track, above the picture, so they can be switched, replaced and exported for each deliverable. Chapter 32 gives the text-engine setting, the caption formats and the export modes.
Never reverse characters by hand to fix their order. Fix a wrong order in the string itself: its direction isolation (the invisible marks that tell a player where a right-to-left run starts and ends), the font, the numerals. Arabic needs correct direction, joining, punctuation, numerals, line order and font shaping. Mixed Arabic, English and numerals need an explicit visual order and a check on the device and the file. When a caption reads correctly in Premiere and wrongly on the phone, the check was made on the wrong player: test on the one the audience uses, export again, and test again.
Check captions against the lips in the real player, with Arabic shaping and order right and every protected pause still present. If a transcript's timing is unusable in dialect, mark the caption timings by hand from the waveform.
Arabic in the frame: always exact
Arabic can enter the picture in two ways: as editorial text composited in post, and as lettering inside a generated frame (a sign, a label, a screen). Both meet the same standard.
Arabic text in a frame must match the approved string letter for letter. A final ى where ي was asked fails; so does a missing dot, a dropped letter, a wrong joining form or a different numeral. Brand names, product names and legal lines are never left to the generator: composite them in post from the real asset or the approved string. Where lettering inside a generated frame fails after the model was given the exact string, fix it in post or re-run the shot, and label the change of route.
Check every Arabic run against its source, letter by letter, at 200 % and in the delivery player. Copy the string from the approved source rather than retyping it, compare the frame with it one letter at a time, and count the words. A caption or title that passes on the editing monitor and fails in the player is a failure.
30.8 Aspect ratios and versions
The brief chooses the primary aspect ratio, and the production is composed natively for it. Every other placement is planned, not assumed. Safe reframing is done in the edit, but a hero shot is regenerated in the secondary ratio when cropping destroys its subject scale, geography, text or brand geometry. There is no universal platform-safe rectangle.
Figure 30.4 — Classifying a shot for a secondary ratio. Manual reframing comes before automatic, and a crop that destroys a shot's meaning sends the shot back to be made again, not to be squeezed.
Planning versions
Overlay every target ratio on the frame upstream, before the shot is made, so that composition can serve all of them.
Classify each shot: crop-safe, needing a reframe, or needing separate origination.
Reframe by hand before you try automatic tracking, and disable automatic subject switching where two faces compete.
Review faces, hands, product, copy, captions and interface overlap over time, across the whole shot, not at its first and last frames.
Recompose or regenerate separately when a crop fails, routing the shot back to the workflow that owns it.
Build cutdowns from story obligations, and quality-check every version independently.
Write each version a manifest line and checksum it.
Overlay first. A frame composed for one ratio and cropped for another loses what was near its edges. The cheapest fix is the one made before the shot exists.
Manual reframe before automatic. Automatic reframing is not proved and needs testing before it becomes a default. A tracker that jumps between two faces, or crops the product, is a known failure; keyframe the narrative subject and the product region yourself, and cut, or originate the shot separately, where both cannot coexist.
No platform figure is given for safe zones. The pages that publish them were third-party blogs or could not be read (checked 30 Sep 2026). Ask the delivery owner for the overlay and measure your own frames against it.
Build a cutdown from the film's obligations, not by squeezing the full version. For a 15-second or a 6-second cut, choose the minimum beats that carry the story at that length, remove subordinate shots, keep the product, the claim and the dialogue meaning, and retime the music, the sound and the captions. Do not simply speed up the 30-second film.
Each version has a manifest line, its row in the file manifest (31.1). It lists the source timeline, the aspect ratio, the duration, the reframed or regenerated shots, the copy and captions, who owns the safe zone, the transforms, the approvals, the file, its hash and the rollback. A shot originated separately keeps its own run record (8.7).
Research exemplar · food and drink · a vertical reframe brief, 16:9 to 9:16, done in the edit · copied exactly from the July 2026 research; instructional, never run
Example 30.2 — 16:9 to 9:16, protecting the bottle and the hand
Vertical version: track the bottle within the central protected column; retain the complete hand contact; crop background only; move post-owned copy into the approved upper safe zone; do not stretch, mirror, or crop the cap/base.It names what the crop may lose (background) and what it may not (the product and the hand contact).
Post-owned copy moves into the approved upper safe zone instead of being cropped or stretched.
If the hand contact leaves the protected column, the shot is not crop-safe: classify it again (Figure 30.4).
Do not assume a vertical version is a crop of the horizontal one. For a shot with speech or a travelling object, reframe by hand and judge it over its whole length. A shot whose crop fails goes back to generation with its state written into the request (29.11).
30.9 Motion graphics: native or generated
Choose native graphics or generated pixels by what may change later. If the client may change the copy, a number or a claim, or if editable graphics were promised, build them natively; generate them as pixels only when the words are final and baked pixels are acceptable. An MP4 is never an editable project. Editorial text is post-owned (30.7) and composited from approved copy whichever route the graphic takes.
Template: the graphics route decision (written for this book; not run)
Graphic: [name] · copy may change after delivery: [yes/no] · data may change: [yes/no] · editable graphics promised: [yes/no] → route: [native / generated / composite] · reason: [one line] · builder: [motion designer / MOGRT]Native graphics
Who builds them. After Effects or Premiere motion graphics templates (MOGRTs), built by a human motion designer. An assistant plans and briefs; it does not build After Effects projects unless the director adopts a route for that. A vendor's own motion-design assistant built into a generator's platform is not used.
What goes in. The purpose, the audience, one takeaway, the call to action, the duration and format, the exact copy, the references with their roles, truthful chart geometry, and the product's state logic: what state each element is in, and what changes it. A planning brief to the assistant ends with "stop after the plan".
Building a native graphic
Plan. A concept, a separate copy list, timed beats (framing, dominant action, entry and exit, reading time, a sound cue for each), a contact sheet and an asset inventory. The director approves.
Split per element. Text, shapes, interface labels and paths built natively; material-rich art as separate images.
Build. The static layout first, corrected against the artwork; one camera controller plus local animation; a full motion pass before any polish.
Inspect. A hero frame, the key transition, the full playback, the loop boundary, the wording letter by letter; retype a longer title to test that the text is editable.
Revise by beat. Name the version, the protected elements, the time range, the defect and the new behaviour; give old and new wording for a copy change; re-render the beat and its neighbours as a new version.
Give each reference its own authority: the artwork controls layout and style, the storyboard framing and motion, the written shot plan the exact lettering.
Let related elements arrive in a readable order (container, text, annotation, reaction), one dominant action at a time.
Put a sound's transient on the frame of visible contact, not on the start of the entrance (29.9).
Protect the product's state logic: "ready" is not "done".
The delivery of a native graphic is the project, its source assets, the stems, an MP4 and the list of checks actually performed, with the graphic's route and its reason.
Generated motion graphics
Motion design can be generated as pixels on Seedance 2.5 (17.3), when no later change of copy or data is expected.
Build the film around one connecting mechanism, something that carries the time from phase to phase, such as a travelling band of light, not stacked effects.
Map every phase to what it proves, and fit the copy to real reading time. A decorative chart that proves nothing is an invented claim; charts need truthful units and geometry.
Replace every source brand, reference and setting with the campaign's own facts.
The prompt carries the duration, the phases with their times, the connecting mechanism, what each reference gives and must not give, and the sound; duration, aspect and the sound switch are set on the model. The timestamps in the prompt are pacing, not commands.
Composite exact copy natively, and label the change of route when generated lettering fails. Generated type is pixels, and the two-layer rule of 10.12 still holds.
Inspect generated graphics in motion: text letter by letter, the product's proportions through the move, and the ending. Wrong lettering goes to a native composite; an expected change of copy sends the whole graphic to the native route. Deliver the clip with its run record, and a note if any text was composited.
30.10 Which tool for which job, and the gate
Job | Tool | Where |
Conform, colour, captions, graphics, versions, final export | Premiere Pro | |
Upscale or frame-rate boost | Topaz Video, the Topaz panel in Premiere, or an upscaler on the platform | 32.6, 32.9 |
Generative Extend (a few frames of handle) | Premiere Pro | 32.5, 32.9 |
Replacement voice, timed dialogue, sound effects | ElevenLabs | |
Music cues | Suno, Eleven Music | Chapters 26 and 27 |
Job numbers, the ledger, the run record | Higgsfield | |
Money, rights and terms | the tool's terms and the client contract |
The records this stage writes. The probe record of every file; the finishing log (each transform with its tool, version, settings and range); the two upscale previews and the director's choice; the colour interpretation, look and output transforms; the loudness measurement beside each final export; the caption files with their approved text; the manifest line of every version.
Before the finish is signed
The picture was locked, and accepted, before any finishing money was spent.
Every file probed; rate and raster conformed on purpose, nothing stretched.
Defects repaired by cut, trim, bridge or child shot before any global transform.
Each transformation judged alone, in motion, on a difficult interval; any upscale compared with the native version at the same grade; generating at the target size priced first.
No identity or lettering changed; the native original kept beside every labelled copy.
Colour: the source interpreted and recorded, no camera-log transform, shots matched to the masters, the look applied separately, viewed through each output transform.
Texture added last, at final resolution, and never asked of a generator.
Unchosen native tracks stripped; one bed per scene; loudness, true peak and downmix measured against a named specification.
Captions from approved copy, timed to the audio that shipped, checked in the real player; every Arabic run in the frame compared letter by letter with its source.
Every target overlaid, every shot classified, every version with a manifest line and its own quality check.
Every graphic's route (native, generated or composite) recorded with its reason.
Open questions
Upscale before or after colour, for generated footage. No maker states an order; only a comparison on your own footage settles it. Until then, upscale from neutral footage and let the comparison overrule.
Upscale after lip sync, or a plate made at the delivery size. Test both on the hard interval.
Loudness figures for Egyptian, Saudi and Gulf broadcasters, cinema advertising and the social platforms. None found; ask in writing.
Safe-zone figures for any platform. None read from a publisher; ask for the overlay.
What to remember
Lock the picture first; nothing is finished before then, and only the intervals the cut uses are finished.
The order: conform, repair, upscale only if needed, colour, then texture last. A comparison on the real interval overrules the default.
Probe, then conform on purpose; never stretch a raster to hide a mismatch, and never assume a label is true.
Judge each transformation alone, in motion; an upscale is a comparison with the native version at the same grade, and the delivery size decides.
Generated media is not camera log: interpret the source, match to the masters, apply the look as a separate step.
A still's finish block describes capture; delivery grain and halation are added in post, restrained, last, at final resolution, and never asked of a generator.
Measure loudness on the final export against the specification the delivery owner names; where none is named, ask in writing.
Captions: approved copy, timed to the audio that shipped, checked in the real player, never reversed by hand. Arabic in a frame matches the approved string letter for letter.
Overlay every target upstream; classify each shot; reframe by hand before automatic; every version has its own manifest line and check.
Graphics that may change are built natively; generated graphics are pixels, and exact copy is composited.
Comments