AI Video Generation Is Expensive—So I Built a 15-Second Remix Workflow

Generating AI video looks effortless in a demo: write a prompt, wait a minute, and receive a cinematic shot.

In practice, the expensive part is not only the subscription or the credits. It is the uncertainty. A usable five-second clip may require several generations because the face changes, the motion breaks, the composition drifts, or the shot simply does not fit the rest of the sequence.

That led me to a different question:

What if creators did not have to generate every frame from scratch?

I tested a more practical workflow: start with video material I am authorized to use, analyze it for high-impact moments, rebuild the pacing, replace the original audio, and keep the result as an editable timeline instead of a disposable render.

The tool behind the experiment is Timeline Studio, an open-source, browser-based video editor:

👉 GitHub: MartinDelophy/ai-video-editor

The experiment

The input was simple:

• one source video of roughly two minutes;
• one music track of more than five minutes;
• a second video containing a specific three-second ending shot.

The target was a 15-second short-form edit with:

• multiple memory-like fragments;
• a clear emotional build;
• a transition in the middle;
• a different final shot;
• the original source audio removed;
• a musical highlight near 1:20;
• a real ending, rather than an abrupt cutoff.

The final structure looked like this:

Time	Editorial function
0–3s	Establish the character and atmosphere
3–7s	Build a fragmented memory rhythm
Around 7s	Use a short flash transition as an emotional hinge
7–12s	Increase energy while protecting readable hero frames
12–15s	Switch to the requested ending shot and fade to black

The final three seconds came from 2:48–2:51 of the second source. Its embedded audio was muted, so the entire edit used only one clean music track.

Why “highlight detection” is not random trimming

A common automation mistake is to cut a long video into equal intervals and keep the most active sections.

That produces motion, but not necessarily meaning.

For this experiment, candidate moments were evaluated with several kinds of evidence:

• audio-energy change;
• subject-aware motion and frame difference;
• shot-change confidence;
• expression change for a tracked face;
• clarity around the main subject.

A practical scoring model can combine these signals:

highlight_score =
    0.30 * audio_energy_change
  + 0.25 * subject_motion
  + 0.20 * shot_change_confidence
  + 0.15 * expression_change
  + 0.10 * subject_clarity

But the score is only a candidate signal. It should not override continuity, intentional stillness, anticipation, or the most important result frame.

After saliency analysis, each retained beat still needs an editorial role: setup, rise, pre-impact, peak, aftershock, or bridge. Without that hierarchy, every shot receives equal weight and the montage feels flat.

Finding the musical high point

The user requested the highlight near 1:20, so I searched around that region instead of blindly cutting from exactly 80 seconds.

The selected range was approximately 80.5–95.5 seconds. That gave the 15-second edit a more natural musical phrase and enough room for a controlled ending.

A simplified FFmpeg extraction looks like this:

ffmpeg \
  -ss 80.5 \
  -t 15 \
  -i input.aac \
  -af "afade=t=in:st=0:d=0.15,afade=t=out:st=14.3:d=0.7" \
  -c:a aac \
  -b:a 192k \
  music-highlight.m4a

This does four important things:

1. starts close to the requested musical moment;
2. limits the result to exactly 15 seconds;
3. avoids a hard entrance;
4. fades the final 0.7 seconds so the music and picture resolve together.

The loudest passage is not automatically the best passage. A useful music highlight also needs a readable beat, a complete phrase, an emotional lift, and a point where the ending can land.

Creating the “memory fragment” feeling

A memory montage is not just a color preset.

The feeling came from combining several editorial decisions:

1. Short, non-contiguous fragments

The first 12 seconds use eight short clips instead of one continuous excerpt. The clips come from different source times and are reorganized around emotional progression.

2. A restrained flash transition

A brief flash near the middle separates two emotional phases. It is intentionally short and placed away from the decisive frame.

Transitions should clarify structure. If an effect hides the action viewers need to see, it is working against the edit.

3. A protected ending

The requested ending shot occupies exactly 12–15 seconds. The final 0.7 seconds fade to black, giving the piece a deliberate closing gesture.

4. One coherent audio bed

All embedded source audio is muted. Keeping only the selected music highlight avoids doubled sound and makes later remixing predictable.

The workflow stayed editable

A rendered MP4 is only the delivery artifact. The actual source of truth is the editable timeline.

The project contains:

• eight memory fragments;
• one three-second ending clip;
• source-time trims for every segment;
• a middle flash transition;
• a final fade;
• one music segment;
• muted embedded audio;
• normalized output dimensions and frame rate.

This matters because editing is iterative. If the ending needs to start 300 milliseconds earlier, I should not have to reconstruct a giant filter graph or repeat the full analysis.

Timeline Studio stores the edit as a .timeline project, so clips, timing, transitions, audio, and effects can continue to be adjusted visually or through automation.

The screenshot above captures part of the actual production process. The ending source ran at 30 fps, while the main sequence used 29 fps, so the render pipeline normalized the final output and fitted the new shot to the project resolution without reintroducing its original sound.

Technical validation

A video is not finished when the encoder exits successfully.

For the final result, I verified:

• duration: 15.000 seconds;
• resolution: 1906 × 1080;
• frame rate: 29 fps;
• total frames: 435;
• video codec: H.264;
• audio codec: AAC stereo;
• full decode: successful;
• the cut at 12 seconds;
• the fade between 14.3 and 15 seconds;
• no embedded audio leaking from the ending clip.

Basic metadata inspection can be automated with ffprobe:

ffprobe -v error \
  -show_entries format=duration \
  -show_entries stream=index,codec_name,width,height,r_frame_rate,channels \
  -of json output.mp4

I also recommend decoding the complete output once instead of trusting container metadata alone:

ffmpeg -v error -i output.mp4 -f null -

That catches broken frames and stream errors that a successful export message may miss.

Why this can be cheaper than generation-only production

This workflow does not replace generative video. It changes where generation is used.

Generation is excellent when a story requires a shot that does not exist. But when usable footage already exists, repeatedly generating near-duplicates is often a poor use of time and credits.

A hybrid pipeline can be more efficient:

1. reuse authorized footage where it already works;
2. analyze and restructure it;
3. generate only the missing shots;
4. import those shots as editable source media;
5. validate the finished sequence as one coherent piece.

The scarce skill is gradually moving from “Can I generate an image?” to “Can I select, structure, pace, and finish a story?”

A necessary note about copyright

“Remix” does not mean downloading and reposting somebody else’s work.

Use footage that you created, generated, licensed, purchased, or received permission to modify. Check whether the license permits derivatives and commercial use. Keep provenance and watermarks when required, and verify music, likeness, and brand rights separately.

Good second creation adds a new structure, viewpoint, explanation, or experience. Removing a watermark and uploading the same content is not a creative workflow—and may violate both rights and platform policies.

About Timeline Studio

Timeline Studio is my attempt to connect local media analysis, browser-based editing, and deterministic rendering in one workflow.

The project is useful for experiments involving:

• content-aware highlight extraction;
• short-form montage generation;
• music highlight selection;
• editable transitions and effects;
• caption and voice workflows;
• AI-generated footage as source media;
• local-first media processing;
• reusable .timeline projects.

If this direction is useful to you, take a look at the repository:

https://github.com/MartinDelophy/ai-video-editor

Issues, feedback, and contributions are welcome.

Final thought

AI video generation lowers the cost of producing footage, but more footage does not automatically create better stories.

Generation determines what material is available. Editing determines what the audience experiences.

For independent creators, the sustainable advantage may not be generating everything from zero. It may be building a repeatable system that knows what to keep, what to cut, when to peak, and how to end.