Case study
Why your AI video step is slow, and a bigger machine won't fix it
A video team's AI presenter took far longer than the clip itself to cut out. The time was going on frames the finished ad never shows, not on machine power.
Key takeaways
- Measure which step takes the time, and what it spends that time on, before buying a bigger machine.
- In this work much of the cut-out went on frames the finished ad never shows.
- Reusing a finished cut-out when only captions, music or layout change can skip the slow step entirely.
- A shortcut that cuts the time but leaves a visible flaw in the video is not a saving.
- An automatic step should fall back to the old result when it fails, never stop the job.
For a video team, an ad that finishes late means fewer versions to test, and every change of caption style or music means waiting again. When AI video processing is too slow, the obvious answer is a bigger graphics machine, and an expensive one. For one client the time was not going on computing power, and the change reached their finishing step the same day.
What a slow step costs a small team
The cost is repetition. A team finishes an ad, then renders it again with another caption style or other music, and waits the same time over.
The reflex is a bigger graphics machine. Here the decision went the other way: leave the graphics machine that makes the presenter to that job, and run the cut-out on the ordinary processors of the machines that put the ads together.
The two looks the team wanted, and why they were slow
Two common looks shared one hidden job: separating the presenter from a background that was not plain. The team wanted the presenter standing over the real footage, and the presenter small in a corner while full-screen clips play.
Neither was filmed against a plain backdrop. The presenter footage came with its own background, so a small AI model had to separate the figure from it in every frame before it could be laid over anything else.
The earlier version worked, and the edges and hair came out clean. It was also slow: the cut-out took far longer than the clip itself, and an ad with the presenter in the corner took far longer than the same ad full screen.
Why AI video processing was so slow here
The model was doing careful work on frames the finished ad never shows. Our engineer measured before changing anything.
On an idle machine, the model alone took a fraction of a second for every frame. The usual tricks were tried and timed, including running several copies of the model side by side, keeping the processors always busy, and a lighter, less precise model. Each one was slower, not faster.
The turning point came from looking at what an ad actually shows. Each spoken line is made as one take, but an ad plays only part of it, and the cut-out figure appears only while a clip plays behind or beside it. In the test ad, many frames of the spoken takes never appeared on screen, and the figure was drawn over even fewer. Every frame was being cut out anyway.
What we changed
The cut-out now follows the video player's own reading of each line, and sends to the model only the frames where the figure is drawn over a clip, with a small margin on either side. The cut-out file ends at the last frame that is needed. Where the presenter sits small in the corner, the figure is read at a lower resolution, which the model handles more quickly.
Two things protect the rest of the job. A line whose cut-out fails, or that has nothing to stand in front of, plays as it did before, so a failure never stops an ad. And because the team often renders the same ad again with another caption style, other music or another layout while the lines and clips stay the same, each finished cut-out is now kept on the machine, stored under a name tied to the material it came from. A second render reads it back without running the model at all. Kept cut-outs are checked before they are stored, cleared regularly, and capped in total size, the oldest unused going first.
One shortcut was measured and turned down. Skipping alternate frames would have cut the time again, but it leaves a faint second outline when the presenter moves his hands quickly, and that is visible in an ad.
What the video team has now
Both looks are live on the client's platform, available to the video team on their finishing step, and no graphics machine was added. On the test ad with the presenter in the corner, cutting the presenter out and the whole render both became quicker. An ad full screen without a cut-out remained the quicker comparison.
A later measurement on one machine showed where the remaining difference goes: a large part in the cut-out, and a smaller part in drawing the figure. Every other look, and every ad with no cut-out, renders exactly as before.
What you can check yourself
- Ask your team how long one video takes from finished edit to final file, and how often they render the same video again with small changes.
- Ask whoever runs your automatic processing which steps take the most time, measured rather than guessed.
- Before agreeing to a bigger machine, ask whether the slow step works on everything or only on what the customer sees.
- Ask what happens when the automatic step fails on one piece: whether the rest carries on or the whole job stops.
- Ask whether a second render of the same video reuses anything from the earlier one, or starts from nothing.
What to do next
Start with a stopwatch and one question: what is the slow step working on, and does the viewer ever see that work. If the answer is no, the saving costs nothing and needs no hardware. Tracing a slow step to the work it really does, before anyone spends on machines, is the kind of job our engineers take on for clients.
Questions people ask
Why does rendering the same video again take so long?
Often because everything is remade from scratch, including the parts that did not change. In this work the lines and clips stayed the same between renders, so a second render of the same ad could read back the finished cut-outs instead of running the model again. If your team renders again with only a new caption style, music or layout, it is worth asking whether the slow step is being repeated for nothing.
Would a faster graphics card fix a slow AI video step?
Not necessarily. In this case the graphics machine was already the right tool for making the AI presenter, and the cut-out ran on the ordinary processors of the machines that put the ads together. Running several copies at once, keeping the processors fully busy and a lighter model were each measured and were slower, not faster.
Can we skip frames to make the cut-out faster?
You can, and it cuts the time again, but the result can show: skipping alternate frames left a faint second outline when the presenter moved his hands quickly. In this work it was measured and turned down. Skipping work on frames that never appear in the finished ad is a different matter, because nothing changes in what the viewer sees.
What happens if the automatic cut-out fails on one line?
Here the line plays as it did before, without the cut-out, so a failure never stops the whole ad. It is a good question to put to anyone who runs an automatic step for you: a failure on one piece should not hold up the rest of the job.