Chapter 32 · Premiere Pro and the finishing tools

Yasser Ashour

2 hours ago

16 min read

Part 6 · Edit, finish and delivery

The cut is built, judged and finished in Adobe Premiere Pro, with the upscalers and the caption tools that sit beside it. This section covers the application as it stands on 30 Sep 2026, what it reads, the settings that matter for Arabic and for delivery, the finishing tools a director can run on the platform or on their own machine, and their limits, prices and terms.

In this chapter

Premiere 26.5, v27 in beta, and what Premiere's generative tools do and do not do
What Premiere reads: timeline files, transcripts, caption files, scripts
The request forms for the upscalers, and three worked settings: an upscale request, an Arabic and English caption spec, a Generative Extend handle
Settings, checks, failures, limits, prices and rights, and what to re-check when a version ships

The craft is in Chapter 29 (the cut), Chapter 30 (finishing) and Chapter 31 (handing over).

32.1 What it is

Adobe Premiere Pro, release 26.5 of 10 Sep 2026 (checked 30 Sep 2026). It is where the cut is built, judged, finished and exported. The assistant plans and checks; Premiere executes and confirms (29.1). Version 27 is in beta: it requires an Apple silicon Mac (M1 or later) running macOS 15 or later, Intel Macs stay on 26.x, and Adobe has not set a release date. Record the build for each project.

The finishing tools beside it:

Generative Extend adds frames to a clip's end or start inside Premiere. It is Firefly-powered, and Adobe states it is trained only on content it has permission to use.
Captions and Speech to Text. Speech to Text lists eighteen languages and no Arabic; Arabic captions are built from approved copy (30.7).
Topaz Video, desktop version 1.7.1 (18 Sep 2026), and the Topaz panel for Premiere, a free panel that sends clips to Topaz's cloud and returns them on a new track.
Upscalers on the platform (Chapter 33): upscale_video with Topaz or ByteDance as the engine, and fps_boost. The platform's changelog also lists a FLUX 3 Video Upscaler (27 Aug 2026); whether the connector exposes it is not yet known.
FLUX Video Upscale from Black Forest Labs, used directly.

Not used. Premiere's AI Assistant (public beta since 23 Jun 2026) runs "a mix of Firefly models and partner models depending on your request", so no request names its engine. The Generative Media Tool generates video and sound effects inside the timeline, outside the run record and the spend packet (8.7, 8.10). Missing coverage is generated on the platform, not in Premiere.

When to reach for something else. A handle longer than two seconds, or a change of action, is a new generation (Chapters 16 to 18). Voices are Chapter 23. Native graphics beyond captions and titles are built in After Effects by a motion designer (30.9).

32.2 How it reads what you give it

Each input owns one thing.

Media owns the picture and its true rate and size; the probe, not the label, says what they are (30.2). The sequence settings own the timebase and aspect and come from the delivery plan, never from the first clip.
A timeline file owns placement, not effects. Premiere imports Final Cut Pro 7 XML with File > Import; Open Project refuses an .xml. In it in and out select frames from the source and start and end place them in the sequence; the end must equal start + (out − in), or Premiere trusts start and end and the source out-point drifts by a frame. Keyframe times and clip-marker positions are counted in the source clip's own frames, not the sequence's. At a transition Premiere's own export writes −1 as the start or end of the touching clips and extends their source ranges by half the transition; normalise that before comparing files.
Premiere's XML export (File > Export > Final Cut Pro XML, Adobe page of 7 Jan 2026) writes a report called "FCP Translation Results"; Adobe warns that only a few effects and features transfer, and that pan, gain and level may not. Read the report and the imported result. FCPXML (Final Cut Pro 10 and later) is a different format that Adobe has said must be converted before Premiere can import it (page of 2 Apr 2026; current wording unconfirmed); renaming the extension converts nothing. OTIO is not shown to import into Premiere: keep it as a record, not a delivery. EDL and AAF are not evaluated.
Transcripts own word timing only. Speech to Text lists eighteen languages, from Danish to Traditional Chinese, and no Arabic; Adobe's language page is dated 22 Aug 2025, and a request on Adobe's forum of 26 Jan 2026 reports the gap in Premiere 26. Adobe's reply pointed to a JSON transcript import with word-level timecodes, which feeds captions and text-based editing; its schema is a separate specification file that is unread: check it before relying on the import. ElevenLabs' Scribe speech-to-text model lists Arabic with word timestamps and no dialect detail (read 28 Sep 2026): check every timing it gives against the waveform. Paper Edit (Premiere 26.5) and text-based editing work from whichever transcript the project holds.
Caption tracks own caption text and its styling. Each track has its own Text Engine setting (32.6), and caption tracks export as burn-in, embedded or sidecar files.
Scene Edit Detection (Clip > Scene Edit Detection) splits a multi-shot generation (Chapter 13) into cuts, a bin of subclips or markers only, and cuts the linked audio too: unlink the audio first when the clip's sound must run across the new cuts. A detected boundary is a place to look, not an editorial decision.
The scripting interface (UXP), Premiere 25.6 and later, inserts, overwrites, clones and removes clips in transactions and inserts motion-graphics templates. It needs Developer mode, which only the person at the machine can switch on, and reportedly does not expose a clip's speed or a caption's text. It has not been tried on a handoff (31.4).
Generative Extend reads the clip's own frames at the edge you extend. It needs a video clip of at least two seconds (an audio clip of at least three).

32.3 The prompt, in this tool's dialect

Premiere takes no prompt. It takes files, settings and a plan; words go to the assistant, and the assistant's product is a plan, a caption file, a checklist or a request block. Keep those in the forms below.

The order Premiere wants. Sequence settings before media; media probed before it is placed; the Text Engine set on a caption track before the Arabic goes in; captions on their own track above the picture; generated frames and transforms listed as you make them (29.14); the export last, then probed.

The upscale request, by engine. These are the request forms on the platform (contracts read 27 to 28 Sep 2026).

upscale_video with provider: topaz: video_id (the finished job's number), resolution 1080p or 2160p, aspect_ratio auto or a fixed ratio.
upscale_video with provider: bytedance: video_id, width and height (required, from the probe), resolution 1080p, 2k or 4k, fps 24, 30 or 60, preset common, aigc, short_series, ugc or old_film.
fps_boost, through generate_video: the maker must be named, Topaz or ByteDance; fps 16 to 120 (ByteDance from 24); the clip goes in as a video reference.
FLUX Video Upscale: input_video, upscale_factor 1.5 to 3, creativity 0 or 1, an optional prompt.

None of the platform's upscalers reads a prompt, and FLUX's optional prompt only steers detail. Everything else is a setting (32.6).

Transcript-led cleanup, for presenters only. Apply silence and mistake cleanup to a presenter or narration, never to acted dialogue: a pause in a performance is a choice, and zero cuts is a valid result. Keep the source untouched; list the candidate cuts with their context; the director approves; apply only those; time graphics and captions to the cleaned timeline; review every join in the encoded file.

32.4 Templates

Written for this book · finished-grade template · finishing requests for Premiere and the upscalers · template; not run

Finished-grade template — three requests

UPSCALE PREVIEWS · [shot] · interval [in, out) of [file], sequence [start–end]
Probe: [size] · [fps] · [frames] · colour tags [tags] · audio [streams]
Delivery: [size, aspect] for [deliverable]. Grade: [one grade for both previews]. Framing: identical.
A: native original scaled to [delivery size] in Premiere.
B: [engine and settings, exactly as in 32.6] from job [id].
Compare in motion at 100 % and 200 %: [skin / hair / edges / label / Arabic letters / reflections].
Record: engine, version, settings, date, debit. Recommend for the delivery size.

CAPTION SPEC · [film] · version [vN]
Canonical lines: [approved Arabic, copied, never retyped]
Spoken dialect: [dialect]. Subtitle languages: [ar as spoken / en approved translation]
Timing authority: [the locked take / the clip's own audio] · file [name]
Tracks: one caption track per language; Text Engine per track: [setting]
Export: [sidecar / embed / burn-in] per version. QC: in [the real player] on [device].

GENERATIVE EXTEND · [file] · [head / tail] · +[frames] at [fps] (limit 2 s)
Needed because: [operation] needs [n] frames beyond the interval; the file has [m].
Before: probe; note the last clean frame. After: inspect every new frame at 200 %.
Record: the transform line (range, tool, version, date). Fallback: [hard cut / shorter dissolve / longer generation].

Written for this book · test-grade template · one cheap proof of a finishing route · template; not run

Test-grade template — one interval, one setting

Proof of [route] on [one 3–5 s interval of the hardest shot]. One engine, one setting, [source size] to [target]. Native original beside it. Slated "DRAFT — NOT FOR APPROVAL". Look only at [face / label / edges] in motion. Record the debit and the verdict. Not a delivery; never shown as the model of a finished version.

32.5 Examples at the standard

Written for this book · automotive and luxury · upscale request through the platform, Topaz engine, 2160p; one 5-second hero shot of a fictional saloon · not run

Example 32.1 — an upscale request for a 4K hero shot

Hero shot S05, the saloon on a Riyadh boulevard at dusk. Delivery: 3840 × 2160 at 24 fps for the cinema and outdoor versions. The accepted clip is job [id], probed at 1920 × 1080, 24 fps, 121 frames, no audio stream.

Make two previews of interval [12, 108) with the same grade and the same framing.
Preview A: the native clip scaled to 3840 × 2160 in Premiere.
Preview B: upscale_video · provider: topaz · video_id: [id] · resolution: 2160p · aspect_ratio: auto.

Judge each preview in motion at 100 % and 200 %: the Arabic badge on the boot lid letter by letter, the wet-road reflections and headlamp edges for shimmer, the driver's hand on the door, paint highlights for banding or wax. Any change to the badge lettering or the driver's face fails the preview.

Record the engine, the settings, the date and the debit shown in the ledger. Recommend for 3840 × 2160, not by pixel count. If B fails, run the same pair with provider: bytedance (width 1920, height 1080, resolution: 4k, fps: 24, preset: aigc). If both fail, deliver native at 1080p and state the limit. Keep the native original.

The pair differs in one thing only, so anything that shimmers or changes is the upscaler's.
Width, height and frame rate on the fallback come from the probe, and the frame rate stays at the source's because ByteDance's price doubles above 30.
The badge is judged letter by letter: Arabic in a frame is always exact (30.7).
The exit is stated in advance: a second engine, then native with the limit.

Written for this book · telecom and tech · Arabic and English caption spec for a 20-second spot, Saudi dialogue on the exact-take route · Premiere caption tracks · not run

Example 32.2 — a caption spec for Arabic and English

CAPTION SPEC · mobile network launch, 20 s · version v4 · 16:9 and 9:16

Canonical lines (copy from the approved script; never retype):
  L1  «أبشر، الشبكة معك حتى في الصحراء.»
  L2  «سجل الحين، وابدأ اليوم.»
English (approved translation):
  L1  Consider it done. The network is with you, even in the desert.
  L2  Register now. Start today.

Spoken dialect: Saudi Arabic (Najdi), a voice actor's recording, locked take ST_02_T03.
Subtitle languages: ar, as spoken, with no move to Modern Standard Arabic; en.
Timing authority: the locked take, exact-take route. Time each caption to the onset of its first word and the end of its last in that recording; if a line needs two captions, break at its comma.
Tracks: CC-AR and CC-EN, each its own caption track. On CC-AR set Track Settings > Styling Options > Text Engine > South Asian and Middle Eastern before any text goes in. Type or paste nothing into CC-AR from a source other than the approved script.
Direction and shaping: right to left; the full stop stays inside the approved string, at the left end of the run. Never reverse characters by hand. حتى ends in ى; في ends in ي: check both letters. This copy has no numerals and no Latin letters; if a line ever mixes them, set its visual order explicitly and check it in the player.
Brand name, the 5G badge and the legal line: composited as graphics, not in captions.
Export: 16:9 gets sidecar caption files, one per language; 9:16 gets Arabic captions burned in. Each version keeps its own manifest line and its own QC.
QC: play each export in the phone player and the web player the audience uses; compare every Arabic run with the canonical line letter by letter; watch the captions against the lips.

The dialect and the subtitle language are separate fields, and the captions follow the audio that ships.
The text engine is set on the caption track before typing, per track, as Adobe's community guidance gives it; test it on your build.
Direction is set on the track and punctuation stays in the approved string; the result is checked in the real player.
The brand name, the badge and the legal line are left to graphics, where they can be exact.

Written for this book · beauty and personal care · Generative Extend, tail handle for a dissolve on a 5-second serum shot · Premiere 26.5 · not run

Example 32.3 — a Generative Extend handle

GENERATIVE EXTEND · S05_accepted.mov · tail · +6 frames at 24 fps (0.25 s; the limit is 2 s)

Needed because: the cut is a centred 12-frame dissolve from S05 into S06, which takes 6 frames from each side of the cut. S05's interval is [0, 121), the whole file, so it has no frames after frame 120.
Before: probe S05 (1280 × 720, 24 fps, 121 frames). The last clean frame shows the serum bottle at rest on the marble shelf, label toward camera, no hand in frame.
Do: extend the tail by 6 frames only. Extend nothing else.
After: step through frames 121 to 126 at 200 %: the Arabic label letter by letter, the bottle's proportions and cap, the highlight on the glass, the marble veining. Any new hand, any drift of the label, any morph fails the extension.
Record: transform list, "S05 tail +6 frames, Premiere Generative Extend, 26.5, checked 30 Sep 2026; frames 121–126 generated".
If it fails: shorten the dissolve to the clean frames, or cut hard, or ask for a longer S05 next time.

The arithmetic is written first: a dissolve needs 6 frames beyond the interval, and the file has none.
The extension is the smallest that solves the join, far inside the two-second limit.
The extended frames are inspected one by one and listed as generated, so the next person knows which frames are which.
The fallbacks cost nothing.

32.6 Settings

Premiere

Setting	Value to use	Why
Sequence timebase and aspect	from the delivery plan	the first clip may be the wrong size or rate
Text Engine, per caption track	right-click the caption track header > Track Settings > Styling Options > Text Engine > South Asian and Middle Eastern, before typing Arabic	runs the text right to left; an Adobe community manager gave this path on 18 Mar 2021 and an expert repeated it on 7 Dec 2021; captions imported from an external file may need re-creating; test on your build
Right-to-left typing (Essential Graphics)	on for an Arabic title or a two-line caption	puts the full stop and comma at the correct end
Caption export	sidecar file, embedded track or burn-in (Adobe lists all three), chosen per version	each version its own choice and QC
XML import	File > Import	Open Project refuses .xml
Scene Edit Detection	choose cuts, subclips or markers only; unlink audio first when needed	a candidate, not a decision
Match Source (Premiere 26.5)	now takes the source's audio channels and sample rate, downmixes, and warns when a layout is unsupported	read the warning

Upscalers on the platform (contracts read 27 to 28 Sep 2026; all untested, prices unknown until the first ledger line)

Tool	Engine	Settings and defaults	Note
upscale_video, topaz	Topaz	1080p or 2160p (default 1080p); aspect auto	the recommended path
upscale_video, bytedance	ByteDance	width and height required; 1080p, 2k, 4k (default 2k); fps 24, 30, 60 (default 24); preset common (default), aigc, short_series, ugc, old_film	fps above 30 doubles the cost; aigc is the preset for generated footage
fps_boost	Topaz or ByteDance, named	16 to 120 fps	disclose and inspect

The catalogue also lists topaz_video and bytedance_video_upscale with extra text fields and a model version: untested, left alone. Higgsfield's own deflicker, generic video upscaler, reframe and background removal name no engine and are not used.

Topaz on your own machine. Topaz Video 1.7.1 runs Starlight Precise 2.6 locally or in the cloud: it targets realistic faces, skin, materials and text in generated footage, takes input at or under 4K, outputs up to 4K at 2×, 3× or 4×, and has three sharpness levels. Its other families include Proteus, Iris, Nyx, Rhea, Gaia and Theia, and frame interpolation. The Premiere panel offers Proteus, Iris, Nyx, Rhea, Starlight Precise 2.6, Astra 2, Starlight Fast 2 and Starlight Mini, plus Hyperion 2.5 for SDR to HDR, all in the cloud and without real-time previews.

FLUX Video Upscale (Black Forest Labs)

Setting	Values	Default	Use
creativity	0 (precise), 1 (creative)	1	0 for faces and products; 1 invents detail and does not preserve identity as strictly
upscale_factor	1.5 to 3	2	the smallest that reaches the delivery size
prompt	optional	none	steers detail only

32.7 Checks before you sign

An upscaled or boosted clip

Probed against the source: same duration and frame count (unless a boost was ordered), same aspect, colour tags unchanged, audio kept.
Compared with the native original at the same grade, in motion, at 100 % and 200 %: faces, labels and every Arabic letter unchanged; no waxy skin, edge shimmer or lost grain; after a boost, the mouth frame by frame.
Labelled "upscaled (tool, version, date)", with the native original kept.

A Generative Extend

Every new frame inspected at 200 %; no new hands, no drift of label or product; extended audio heard through its splice.
Listed in the transforms with its range, tool and version.

Captions

Every Arabic run compared with the canonical line letter by letter.
Direction, joining, punctuation and numerals right in the real player, on the audience's device.
Timed to the audio that shipped, against the lips.
One file per language per version, each opened.

An import or a final export

The translation log read; first, middle and last frame of each placement seen.
Duration within one frame of the plan; loudness measured against the named specification (30.6).
The export probed and reopened.

32.8 Failures and fixes

You see	Cause	Fix
Arabic caption characters reversed	the caption track's Text Engine is not the South Asian and Middle Eastern one (typical after an SRT import)	set it in Track Settings, re-create the captions, never reverse by hand
Full stop or comma on the wrong end of a second Arabic line	typing direction	switch on right-to-left typing in Essential Graphics
No transcript for Arabic dialogue	Speech to Text lists no Arabic	use the approved script, or import a JSON transcript from another tool; check its timings against the waveform
Open Project refuses the XML	it is an import, not a project	File > Import
Source out-point drifts by a frame	end typed, not computed	end = start + (out − in)
An effect or fade is missing after import	it did not translate	read the translation log; rebuild it natively
Skin waxy, a label or badge changed	creative or generative upscaling	precise mode or a lower factor; failing that, native and composite the text in post
Charged double	ByteDance above 30 fps	keep the source rate
An extension drifts from its clip	Generative Extend invents the frames	shorten it, choose another handle, or generate longer
A panel result cannot be previewed	the Topaz panel has no real-time preview	run a 2-second test first

The ladder for a finishing defect, smallest lever first: a shorter interval or another select (29.12) → a different setting on the same engine → the precise mode or a lower factor → the other engine → deliver native and state the limit → regenerate at the target size.

32.9 Limits, prices and rights

(checked 30 Sep 2026)

Premiere. Single app US22.99amonthontheannualplanbilledmonthly, with25generativecreditsamonth; CreativeCloudProUS34.99 for the first three months and US$69.99 after, with 4,000 a month (US prices). v27 needs macOS 15 or later on Apple silicon; Windows needs Windows 11 24H2 and a GPU.
Generative Extend. Up to two seconds of video and ten seconds of audio; a video clip of at least two seconds and an audio clip of at least three. Since 26.5 it supports 360p to 4K UHD, any aspect ratio, interlaced and non-square-pixel footage; one reviewer reports that it generates at 30 fps at most and outputs SDR 8-bit, so try it on an HDR or 60 fps clip before relying on it. It is a premium feature that spends generative credits; the per-extension cost was not readable.
Speech to Text. Eighteen languages, no Arabic. Captions can be translated into twenty-seven languages; whether Arabic is among them is not yet known.
Topaz. Personal plan US$39 a month (billed monthly over twelve months) with 25 monthly video cloud credits and unlimited local rendering; the Premiere panel is free and uses cloud credits.
Platform upscalers. No quote before the job for upscale_video; the price shows in the ledger afterwards. A typical job takes one to five minutes; failed jobs are refunded. ByteDance above 30 fps doubles the cost. Price generating at the delivery size first (30.3).
FLUX Video Upscale. Source MP4 of at most 20 seconds, 50 MB and 2560 × 1440, at least 480p; output 1080p, 2K or 4K, up to about 14.4 megapixels a frame, at 1.5 to 3 times the source; audio kept. Billed per megapixel-second of output: precise US0.07(aboutUS0.14 a second at 1080p), creative US0.10(aboutUS0.20).
Rights. Adobe states Generative Extend is powered by "commercially safe" Firefly, trained on Adobe Stock and public-domain content and never on users' content. Generated frames go in the transforms list, and a client contract that restricts AI-generated material applies to them (Chapter 34).

32.10 Version notes

Premiere 26.5 (10 Sep 2026): Generative Extend expanded to 360p through 4K UHD and any aspect; Paper Edit from transcripts; the Generative Media Tool; Match Source audio channels. 26.3 (June 2026) added single-word captions. No change to interchange or scripting is listed.
Premiere v27: beta builds since the week of 3 Aug 2026; Apple silicon and macOS 15 required; historic majors shipped November to January. Re-test the XML import, the caption Text Engine and Generative Extend on the first v27 build.
Topaz: Video 1.7.1 (18 Sep 2026) with Hyperion 2.5; the updates page of 28 Sep 2026 announces Starlight Fast 3 and a Speed Boost for Starlight Precise 2.6, "coming soon to Topaz Video". Re-read the model list before a job.
Platform: ByteDance Upscale arrived on 14 Jun 2026; the FLUX 3 Video Upscaler was listed on 27 Aug 2026.

Open questions

The JSON transcript import. Its word-level schema is a separate file that has not been read; one Arabic transcript imported and timed against the waveform would settle whether it serves Arabic captions.
The Text Engine setting. The path comes from Adobe community replies of 2021; confirm it on your build.
Whether the platform exposes the FLUX 3 Video Upscaler. Check the connector's upscale options before planning around it.
Overlaps through the XML. Whether a J or L cut survives the interchange file is not yet known (29.7).
Generative Extend's frame rate and range on HDR and 60 fps clips.