Part IX — Editing, Subtitles, Pacing, and Finishing
Chapter 51. Color, Visual Consistency, and Technical Export#
In this chapter
51.1 Match first, style second#
Material from different models is unified for exposure, white balance, skin tone, black level and saturation before the project's look is applied. Dropping a LUT on top magnifies errors, and it cannot fix a light direction or identity drift.
51.2 Shot matching#
Reference the approved styleframes and the adjacent hero shots. Compare waveform, RGB parade and skin tone first, then judge visually. Reverse angles may differ in lighting, provided the difference comes from the world's sources.
51.3 Skin tone#
Preserve real differences between characters rather than grading everyone to one complexion. Check that face, hands and neck agree, and that composited characters merge with the environment's color temperature.
51.4 Black level and highlights#
Phone screens and platform compression eat shadow detail. In low-light scenes, keep faces and key props readable without lifting black into grey. Avoid clipping windows, screens and skin in the highlights.
51.5 Sharpness, grain and model differences#
Some material is over-sharp and some over-soft. Unify detail frequency and noise, then add very restrained project grain. Grain is not a universal layer for concealing poor quality.
51.6 Flicker and frame rate#
Check brightness, texture, face and background flicker. Deflicker tools can erase motion or skin detail, so work locally. Unify the timeline frame rate to avoid unintended duplicate frames and unnatural motion.
Review frame-rate conversion and slow motion section by section on action shots.
51.7 Motion blur#
Different generation routes can produce different amounts of blur. Slight matching unifies the sense of motion, and blur must never be used to hide a face or text. Keep proof inserts sharp.
51.8 Upscaling and denoising#
Apply only to approved masters and compare faces, text and edges before and after. Upscaling can invent lashes, teeth and texture; denoising can plasticize. Keep the originals.
51.9 The output family#
The master is highest quality with no subtitles, or with a replaceable subtitle layer. Platform versions carry the target ratio, encoding and loudness. Ad versions have their own durations and CTA. Review versions carry timecode, watermark and version information. Archive versions hold project files, separate tracks, graphics and subtitles.
deliverable:
id: MASTER_E001_V12
type: clean_master
resolution: 1080x1920
frame_rate: project_native
audio: stereo_master_plus_stems
subtitles: separate
color_space: documented_project_standard
checksum: recorded
51.10 Encoding and platform verification#
Export to the platform's current specification, then upload and download or preview the actual transcode, checking shadows, subtitles, sound, first frame and sync. Specifications change; verify the official requirements at delivery.
Never treat one platform file as the long-term master.
51.11 Technical QC#
Check resolution, aspect ratio, frame rate, color, black frames, frozen frames, sync, audio channels, subtitles, loudness, true peak, filenames, versions and hashes. Play it through once in full rather than sampling frames.
51.12 Build a chain of scene references#
Choose one or two hero reference shots per scene and fix skin tone, background neutrality, black level, highlights and look strength there. Match the other shots to the reference, then check the adjacent cuts. Do not let each shot look attractive on its own, because continuous playback exposes jumps in warmth, brightness and texture.
The reference must be an approved shot with correct identity. Using a well-composed generation with drifting skin tone as the grading baseline spreads the error across the whole scene. Scene references also link to the story's lighting state: before the blackout, during it, and after emergency lighting are three explicable states and should not be unified into one exposure.
51.13 Normalizing different generation routes#
Identify each route's signature first: one model's highlights lean cyan, another over-smooths skin, one composite layer lifts black, an upscaler adds hard edges. Build a starting correction per route, and still judge each shot. Put technical normalization before the look node and keep the processing chain switchable.
Match from large to small: exposure and white balance, contrast and black level, skin tone, saturation, local lighting, then sharpness and noise. If the direction of light on a character is completely wrong, grading must not pretend to match — return to compositing or generation.
51.14 Picture acceptance under vertical framing and compression#
Faces occupy a large share of a vertical frame, so small differences in skin tone and texture are more visible, while subtitles, platform UI and heavy compression change local contrast. Grading cannot be completed only on a full-screen professional monitor; check at target size, normal brightness and typical ambient light.
Upload tests focus on shadow banding, gradient stepping, red emergency light, fine hair, grain, fast motion and subtitle edges. Platform encoding can turn fine grain into noise, or make repaired shadows reappear. The platform preview is part of the delivery target, not a scaled-down copy of the master.
51.15 The output matrix and an immutable master#
Produce a QC'd clean high-quality master first, then derive platform, subtitle, advertising, review and archive versions from it. Record cropping, encoding, subtitles, loudness, head and tail, and watermarks in each derivation, and never build other versions from a compressed platform file.
output_profile:
id: OUT_PLATFORM_A_VERTICAL_V06
parent: MASTER_E001_V12
video_codec: project_delivery_setting
raster: 1080x1920
frame_rate: inherit_master
subtitle_mode: burned_zh_cn
ui_safe_template: SAFE_PLATFORM_A_2026Q3
audio_master: MIX_E001_PLATFORM_A_V03
verification: upload_preview_required
Specifications are time-sensitive data and stay separate from the work's master. When a platform updates, change the output profile rather than redefining the work itself.
51.16 Matching the blackout scene#
The blackout scene came from three generation routes: Lin Yun's close-up leaned cyan, the phone insert had a lifted black level, and the crowd wide over-saturated the red emergency light. The team used the shot of Lin Yun after the emergency lighting comes up as the scene reference, unifying the red direction and the background black. The phone screen kept its higher brightness with pure black reduced to match the glass reflection. The crowd's red was kept only on lit edges rather than tinting the whole frame.
After the first platform upload, shadows banded and hair edges flickered. The master was left unchanged; grain and encoding headroom were adjusted on the platform derivative alone and re-uploaded for verification. That protects the high-quality archive while acknowledging the limits of the real viewing chain.
51.17 SOP for grading and export#
First, lock picture. Second, unify exposure, white balance and skin tone. Third, match black level, saturation, sharpness and grain. Fourth, handle flicker, frame rate and motion blur. Fifth, apply the look and check on a phone. Sixth, output the master and the version family. Seventh, run automated technical checks. Eighth, play through in full. Ninth, re-check after platform transcoding and archive the hashes.
51.18 Fault tree#
Symptom: it is less consistent after the LUT. Base matching was never completed. Pull the look layer and correct the shots first.
Symptom: shadows are pure black on a phone. Review happened only on a bright monitor. Test on the target device and through platform transcoding.
Symptom: faces change after upscaling. The algorithm is inventing structure. Reduce the strength or keep the original resolution.
Symptom: subtitles are mushy and audio distorts after upload. Platform transcoding was never re-checked. Inspect the actual published file.
51.19 Checklist, exercises and deliverables#
Check that matching precedes styling; that skin tone is natural; that shadows and highlights are readable; that flicker is repaired; that frame rate is unified; that the master is clean; that platform specifications are verified for the current period; that uploaded transcodes are re-checked; and that the archive carries hashes.
Exercise one: match shots from three different models in one scene. Exercise two: compare faces and text before and after upscaling. Exercise three: design the master, platform, advertising and review output matrix.
Deliverables for this chapter: the color match, the look pass, the flicker report, technical QC, the master family, platform verification, and the archive manifest.
51.20 Match scenes with anchor shots and shot families#
Choose one to three anchors per scene: neutral protagonist skin tone, a representative wide, and any special state. Group the remaining shots into families by camera zone, lighting state and generation route, match within the family first, then unify against the scene anchor. Grading shot by shot from the top lets the correction direction drift continuously.
The matching order is exposure, white balance, black and white points, skin tone, saturation, local light and texture, and only then the style look. If light direction, skin tone category or background structure is wrong, grading should report the upstream problem rather than faking a repair with extreme local nodes.
Approved anchors store reference frames, color space, viewing conditions and version. A replacement shot matches the anchor first and then enters full-scene regression, which prevents a new shot that looks fine alone from jumping when placed back on the timeline.
51.21 Temporal consistency of flicker, texture and sharpness#
AI material often shows nothing on a single frame while skin texture, background detail, exposure and sharpness breathe during playback. Temporal QC uses difference views, waveforms and normal-speed viewing together. Technical tools locate anomalous frames; humans judge whether they are visible and whether they break attention.
Repair from source generation, stabilization and local compositing first, and only then reach for denoising or blurring a whole section. Excessive denoising waxes faces; excessive sharpening makes texture jitter frame to frame. Grain can unify the feel of different models and cannot cover structural drift.
Frame-rate conversion and interpolation enter temporal review too. Duplicate frames cause judder, and bad interpolation fuses fingers and mouths. Motion blur should match speed. Any processing that promises smoothness is re-checked frame by frame at action and facial peaks.
51.22 Immutable masters and derived outputs#
A finished master stores the highest-quality picture, final audio, subtitles and graphics, in an explicit color space, frame rate, resolution and channel layout, bound to no platform's compression. Platform versions, censored versions, textless versions, international versions and acquisition crops all derive from the master or from a defined mezzanine.
Each output records the source master hash, encoder, parameters, cropping, subtitles, audio version and creation time. When platform requirements change, re-derive rather than transcoding an already compressed platform file. The release candidate manifest lists every file and its checksum, so delivery and archiving can be verified.
After upload, re-check orientation, cropping, color shift, sync, subtitles and sharpness online. A correct file before publishing does not guarantee correct playback on the platform; the online verification result is part of the release evidence.
A note on sources#
Visual unity in a finished piece cannot rest on a single filter. Generated material must complete technical matching first, then pass style, output and platform transcoding verification.