Remastering a Legacy Ad: After Effects vs. AI Video Synthesis

Why a legacy commercial is worth rebuilding instead of reshooting

A commercial that aired years ago is not a dead asset. It is a compact record of how a brand once spoke: the casting, the wardrobe, the typeface, the cut rhythm, the music hook that people still hum. When a brand wants to reuse that spot today, the instinct is often to reshoot. Reshooting is usually the most expensive of the available options, and it rarely reproduces the specific charm that made the original work. Remastering keeps the performance and rebuilds everything around it.

There are also practical reasons that push teams toward restoration rather than replacement. Broadcast and streaming specifications have moved on. A spot delivered for a 1080i broadcast chain now has to survive in a 4K feed, on a vertical phone screen, and inside a muted autoplay environment where burned-in text at the bottom of a wide frame is effectively invisible. Audio delivery targets have shifted too, so a mix that once sounded punchy on television can feel thin after platform loudness normalization.

The creative case is just as strong. Nostalgia marketing works because recognition is fast. A viewer who half-remembers a campaign gets a small jolt of familiarity, and that jolt buys a few extra seconds of attention. If the remaster looks muddy and the lettering is illegible, the recognition arrives attached to disappointment. If it looks crisp and deliberate, the same footage reads as a confident callback rather than an old file pulled from an archive drive.

This guide walks through both routes for that kind of rebuild. One route is a compositing-led remaster in a layer-based tool such as After Effects, with Resolve Fusion as an equally valid alternative. The other is a generative route built on AI video synthesis. In practice, almost every serious project ends up as a hybrid, and knowing where the boundary sits is the difference between a clean job and three weeks of fighting artifacts.

Two philosophies: deterministic layers versus generative synthesis

Before comparing tools, compare the underlying assumptions. The two approaches disagree about what a video frame is, and that disagreement explains nearly every practical difference downstream.

The deterministic approach

In a compositing application, a frame is the result of an explicit instruction set. Layer A sits above Layer B, a mask excludes the left third, a keyframe interpolates opacity over twelve frames, an effect chain applies blur and then grain. Given the same project file and the same source media, you get the same output, every time, on any machine. That reproducibility is why advertising, broadcast, and title design still run on layer-based compositing.

The cost is human time. Rotoscoping a moving subject frame by frame is slow. Tracking a logo onto a shirt that wrinkles and folds is slow. Building a matte around flyaway hair is slow. Every minute of finished footage can represent hours of manual work, and revisions multiply that work rather than replacing it.

The generative approach

AI video synthesis treats a frame differently. A model is asked to produce a plausible continuation, restyle, or restaging of what it sees. Instead of instructing, you describe, condition, and sample. You supply a first frame, a depth pass, a pose skeleton, or a text prompt, and the model invents the pixels that satisfy those constraints.

This is enormously powerful for material that does not exist yet. You can extend a camera move that was cut short in the original, generate an insert shot of a product on a table, or create a background plate for a scene that was shot against a crumpled cyclorama. The weakness is exactness. Generative models are good at plausible and bad at specific. They will happily invent the wrong logo, drift a face across a cut, or change the texture of a jacket between shots because nothing in the model knows your continuity script.

Where the two overlap

Most of the AI tools a working editor touches are not generative at all in the dramatic sense. They are assistive. Matte extraction, denoising, upscaling, frame interpolation, and audio source separation are all model-driven, and all of them slot quietly into a deterministic pipeline where they replace a tedious manual step. A defensible rule of thumb: use models for tasks where plausible is good enough, and keep deterministic control wherever a client, a legal team, or a continuity supervisor will inspect the result.

The pre-flight audit: what to settle before opening any application

Remasters fail in the first hour far more often than in the last. Spend that hour on paperwork and inventory.

Inventory the source material

Find every version of the spot you can. Camera originals or online masters are ideal, but a surprising amount of value hides in side files: a ProRes export from the edit, a layered design file for the end card, the original font folder, a LUT, a stem export of the music. Each of these saves hours later. If the only surviving copy is a compressed web upload, accept that your ceiling is lower and plan the look around that limit rather than pretending otherwise.

Map the cadence and format quirks

Write down the frame rate, whether the source is interlaced, and whether speed ramps were built with frame blending or optical flow. Interlaced footage deinterlaced after upscaling looks worse than deinterlaced before. A 23.976 source conformed to a 25 fps timeline will need care or it will drift against the audio. These details are boring and they determine whether your first export syncs.

Check the rights

A remaster is a new use. Music that was licensed for a limited broadcast window may not clear for streaming or social. Talent releases often specify term and territory. Stock footage licenses may forbid redistribution in a new edit. Font licenses frequently cover print and not broadcast embedding. Sorting this early is unglamorous and it is the only way to avoid a finished master you cannot publish.

Define the deliverable matrix

Decide up front how many aspect ratios you owe. A typical modern set is a 16:9 master, a 9:16 vertical cut, and a 1:1 or 4:5 social cut, each with its own safe areas for text. Vertical crops of a wide spot usually destroy the composition, which means the vertical version needs its own reframe decisions rather than a center crop.

Workflow A: a compositing-led remaster in After Effects

This is the route to take when the source is decent, the changes are mostly graphical, and the client cares about frame-accurate fidelity.

Step 1: Conform and stabilize

Import the best source, set the correct interpretation, and confirm sync against the audio stem. If the camera drifted, stabilize before you roto; stabilizing after you roto means redoing the roto. For shots with parallax, prefer planar tracking over point tracking, since a plane survives perspective changes that a single tracker point cannot.

Step 2: Isolate the elements you intend to change

Separate the shot into components: subject, background, product, graphics, and anything that must move independently. Assisted selection tools handle clean edges quickly and fail predictably on hair, glass, and motion blur. Plan to clean up the last five percent by hand. That last five percent is what separates a professional matte from an obviously auto-generated one.

Step 3: Rebuild typography as live text

Never scale up a burned-in lower third. Recreate it as vector text with the original typeface, matched tracking and weight, and rebuild any animation with new keyframes. This single decision makes the whole project reusable: you can localize language, resize for vertical, and update a legal disclaimer without touching a single pixel of footage.

Step 4: Clean the plate

Remove dust, scratches, dead pixels, and stray logos. Compression blocking around high-contrast edges will need denoising, and denoising always costs detail, so apply it locally through masks instead of across the whole frame. Where a logo must disappear over moving fabric, expect to combine tracking, patching, and manual paint work.

Step 5: Grade, then texture

Grade the restored footage toward a consistent look, then add grain last. A useful trick is to extract a grain plate from an untouched section of the original and lay it over the restored material so the finished image shares the same texture. Grain applied before a final upscale gets magnified and reads as noise.

Step 6: Rebuild the audio

Split the mix into dialogue, music, and effects. Clean dialogue with spectral repair, replace effects that sound dated, and decide whether the music bed survives. Where a music license cannot be renewed, a re-record or a licensed soundalike keeps the emotional shape of the spot without the legal exposure.

Workflow B: a hybrid AI-assisted remaster

This route shines when footage is missing, the source is badly degraded, or the brief asks for something the original never contained.

Restore the plate first, generate later

Upscale and denoise before any generative step. A model asked to invent motion from a blocky, aliased source will invent around the artifacts and lock them in. Clean the plate, correct the color, and export a high-quality intermediate. Frame interpolation can raise a 24 fps source to a higher rate, but use it sparingly on live action; it produces a distinctive smoothing that reads as artificial on human movement.

Generate only what you cannot find

Use image-to-video for inserts and backgrounds, conditioning on a still you composed yourself. If you need a camera move that the original cut short, generate an extension and blend it with an optical-flow transition rather than a straight cut. For product shots, generate the environment and composite the real product on top; generated product geometry almost always has subtle wrongness that viewers notice without being able to name.

Replace or extend performance carefully

Video-to-video restyling and lip-sync tools can change a performance, but they raise consent and likeness questions that a compositing job does not. Keep written permission for any synthetic alteration of a person, and keep a documented record of what was changed. Beyond ethics, it is practical risk management.

Composite generated material with discipline

Generated shots fail integration for the same handful of reasons: wrong grain, missing motion blur, perfectly clean edges, and no lens imperfection. Add chromatic aberration, a touch of lens breathing, subtle vignetting, and matched grain. Then watch the shot at quarter size and at full size. Artifacts that vanish when small often scream when large.

Quality control: the checks that catch most failures

Run the same checklist on every version before delivery.

• Continuity across cuts: does the subject wardrobe, hair, and prop placement hold between shots?
• Motion coherence: do hands and limbs move with believable weight, or do they slide?
• Text legibility: is every line readable in the smallest delivered format, with the sound off?
• Grain and noise consistency: does the texture change at every cut?
• Edge quality: zoom to 200 percent and inspect mattes against bright and dark backgrounds.
• Audio: check loudness targets, sync at the head and tail, and mono compatibility.
• Color: confirm the grade holds on a phone screen, not just on a calibrated monitor.

Choosing your route: a practical decision table

Situation	Better starting point	Why
Clean master, graphics need updating	Compositing	Deterministic work is exact and fast to revise
Heavy degradation, no clean master	Hybrid AI	Restoration models recover more than manual denoise
Missing shots or coverage	Hybrid AI	Generation fills gaps that editing cannot
Client needs frame-exact approval	Compositing-first	Reproducible output supports review cycles
Many aspect ratios and language versions	Compositing-first	Live text and layered comps scale cleanly
Tight schedule, acceptable creative latitude	Hybrid AI	Generation compresses the slowest manual steps

A note on hybrid planning: front-load the deterministic work. Build the timeline, conform the audio, and lock the graphics first. Then drop generated shots into the locked structure. Doing it the other way around means recutting the generated material every time the edit shifts, which is the most common way these projects go over schedule.

Common mistakes that ruin a remaster

Upscaling before cleaning. Upscalers amplify noise and block artifacts along with detail. Denoise and correct first, then upscale.

Rebuilding everything at once. Change one variable per review pass. Cut structure, then graphics, then grade, then audio. Clients approve faster when they can see what changed.

Ignoring the small screen. A composition that feels grand on a monitor can become a thumbnail where nothing is readable. Test the vertical cut early, not the day before delivery.

Over-smoothing faces. Aggressive denoising turns skin into plastic. Protect faces with masks and leave a little texture.

Treating generated shots as final. They are drafts until they are composited, graded, and textured. Budget for that finishing work or you will deliver something that looks like a demo.

Forgetting the archive. Save project files, LUTs, fonts, and the restored plate. The next remaster will be cheaper because of the work you did on this one.

Where AI genuinely beats manual work

It is worth naming the specific tasks where models win outright, because this is where the hybrid approach earns its place.

Restoration of damaged footage is the clearest case. Models trained on degradation patterns recover detail from scans, old tape transfers, and heavily compressed uploads far beyond what a manual denoise and sharpen chain achieves. Matte extraction on complex subjects is a close second; the assisted selection in modern compositing tools handles edges that once took a full day of frame-by-frame refinement. Frame interpolation, audio source separation, speech transcription for subtitle generation, and automatic reframing for vertical delivery round out the list.

Where models do not win: precise graphic animation, timing-sensitive cuts, brand-accurate product rendering, and anything that must be identical on every render. Keep those deterministic. The practical skill in a remaster is not choosing a side, it is sorting the shot list into two columns and working both columns in the right order.

Frequently asked questions

Can I remaster a spot that only exists as a compressed web upload?

Yes, with realistic expectations. Restoration and upscaling can improve the viewing experience considerably, but you cannot recover detail that was never encoded. Prioritize color correction, grain management, and clean typography, since those three changes do more visible work than resolution alone.

Is a generative remaster cheaper than a compositing remaster?

Not automatically. Generation compresses some expensive steps and adds new ones: curation, integration, and quality control on shots that must match a locked edit. The real savings appear when footage is missing or badly degraded. When the source is clean and the changes are graphical, deterministic work is often the faster route.

How do I keep a subject consistent across generated shots?

Condition every shot on the same reference stills, keep the camera angle and lighting description identical across prompts, and avoid generating a wide range of framings in one pass. Then lock the chosen shots and stop regenerating. Consistency comes from a small, frozen reference set, not from many attempts.

What resolution should the master be?

Deliver the highest resolution your source can legitimately support after restoration, and keep the master as a mezzanine-quality file such as ProRes or DNxHR rather than a compressed consumer format. Downstream platforms will re-encode anyway; a clean intermediate gives them something worth encoding.

Do I need to redo the music?

Only if the license does not cover the new use. Check the term, territory, and media defined in the original agreement. If the music cannot be renewed, a re-record or a licensed alternative preserves the intent of the spot while removing the risk.

How long should a remaster take?

A graphics-and-grade refresh of a clean thirty-second spot can be a few days of focused work. A full restoration with missing coverage and multiple aspect ratios is a multi-week project. The variable that moves the schedule most is not tooling, it is how much of the original material survived.

A closing workflow you can repeat

Treat every remaster as three phases in a fixed order. First, audit: inventory the source, confirm rights, and define the deliverables. Second, restore: conform, clean, denoise, and correct before anything generative touches the footage. Third, rebuild: recreate graphics as live elements, replace or generate only the shots that are genuinely missing, and composite everything back into a locked timeline.

Within that frame, the choice between a layer-based compositing approach and AI video synthesis stops being an ideological argument. It becomes an assignment decision you make shot by shot. Deterministic tools keep the parts that must be exact; generative tools cover the parts that never existed. The remasters that look effortless are the ones where that division was decided at the start rather than discovered in the final week.