Get the free kit
EP3 · making an ad remade · measured 2026-10-11

I remade one ad three times with AI, and here is what finally worked

I gave an agent a twenty-second shoe ad to work from. The second attempt passed its own forty-two-frame check and still had seven problems I only found by looking. The third looked usable and used fewer Vidu credits. This is the whole process, the six checkpoints that got me there, and the numbers, including what the accepted remake still gets wrong.

  • 3attempts at one ad
  • 4,000Vidu credits, all three
  • ≈21 sthe accepted remake

In short: I used an agent called Muse to run Vidu, a video generator, and remake one shoe ad three times. The first attempt used 1,680 Vidu credits and came back with a third shoe, white shoes that were never in the brief, and a box carrying another brand's lettering. The second used 1,280 and passed an automated check of forty-two frames; seven problems were still visible to the eye. The third used 1,040, looks usable, and keeps two known differences from the reference. The lesson that carries over is simple: look at the starting image before you let it animate.

What this was, and what it was not

The goal was to follow a reference ad's shots and pacing, using a different person and replacement shoes. It was a production test, not a client commission, and it was not a test of whether the ad would sell more shoes. I did not compare different tools against each other, and both the product images and the generation mode changed between attempts, so this is the record of one project, not a controlled test.

On the reference itself: the shoes and the reference footage are ours or used with permission, and no other brand is named in this write-up. Where an attempt invented a brand's lettering, that is treated as an error to fix, not a feature.

Round one: 1,680 credits, and a third shoe

Muse helped me choose a person, then generated the shots. What came back had problems on almost every one. The product box picked up another brand's lettering. White shoes came back into the picture even though the brief called for a single black product. One shot had a third shoe floating beside the person. The person changed between shots as well.

And at one point the same video came back under a new filename. Changing the filename had not changed the video. The generation had finished, but I could not use the result. The brief had described what I wanted; the pictures showed what I was actually getting.

Round two: the check that passed everything, and the seven things it missed

I tightened the instructions and tried again. The second attempt used 1,280 Vidu credits. Muse checked forty-two frames from the finished video and gave every one a PASS. Then I watched it myself.

  • The scene changed.
  • There was an extra leg.
  • Something drifted down through the shot.
  • The shoe gained three stripes.
  • A shoe that should have been worn became an empty shoe on the ground.
  • The person changed.
  • An extra black object appeared.

Seven problems in the version the agent had passed. An automated check can tell you that frames exist and are not obviously broken; it does not tell you that a frame is the picture you asked for. Its self-check can help me decide what to inspect, but I still have to look at the actual pictures.

Round three: the job split into six checkpoints

For the third attempt I broke the work into six stages, each giving me something concrete to approve before the next one started.

  1. Break down the referenceFor each shot, record the scene, framing, camera angle, lighting, movement, and the number of shoes.
  2. Write a shot cardSpell out what belongs in the picture, and what would make it a failure.
  3. Make a clean starting image, and check itBuild a still starting image for each shot and compare the person, product and framing against the reference before any video is generated.
  4. Test the difficult shots firstTry the hardest shots before committing to the rest, so a problem shows up while it is still cheap to fix.
  5. Generate the remaining shotsUse the approach that passed the test for the shots that are left.
  6. Compare the finished edit with the referenceCheck it at half-second intervals and around the tricky moments.

Three things the third attempt taught me

  • Check the starting image before you animate it. The single step that saved the most credits was refusing to generate video from a still I had not looked at.
  • A covered-up area can be read as a real object. In one shot the model read a blacked-out shoe area as a loose shoe lying on the grass. Replacing that input with a clean, finished starting image removed the problem.
  • The reference mode we used did not lock the first frame. The image guided the subject; it did not reliably become frame one. Switching to image-to-video with a still I had already checked is what held the shot. For the turn, start-and-end images guided the direction but still needed a visual check, so I kept only a short part of that shot in the final edit.

The version I accepted, and what it still gets wrong

The person and the black-and-red shoes are much more consistent than in the earlier attempts, and the accepted remake runs about twenty-one seconds. Two differences from the reference remain, and I am keeping them visible rather than editing around them.

  • The first shot still holds the box at the waist, lower than the reference.
  • The turning shot ends early with an edit, where the reference shows the full turn.

The credits, and what they do not cover

  • Round one used 1,680 Vidu credits, round two 1,280, round three 1,040. All three together used 4,000. The third attempt used fewer credits than the first two.
  • Credits are not the same as total cost. These figures do not include my time or any other production cost, and credits are not dollars. The product images and the generation mode changed between attempts as well, so the per-round numbers describe what happened, not a fixed price.

Where these figures come from

So you can check them rather than take my word for it.

  • 1,680 / 1,280 / 1,040 Vidu credits are the per-attempt figures from the agent's own reports and the project's recordings. The three add up to 4,000. They do not include human time or other costs.
  • Forty-two frames, all PASS is the agent's own check of the second attempt. The seven problems are from reviewing that same video by hand.
  • About 21 seconds is the length of the accepted third-attempt remake.
  • This page is the record of one project. Your numbers will differ, and a different product or reference will bring different problems.

Made with an agent (Muse) driving a video generator (Vidu). The remake footage is AI-generated. The reference and product footage are ours or used with permission, and no other brand is named. Not affiliated with or sponsored by Vidu or any other product named on this site.

Want the free kit instead?

The kits with the prompts and scripts from the GPU episodes are on the kits page, and the small single-purpose tools for your agent are on the skills page.

The step worth borrowing is free: check the starting image before you animate it. Want a remake for your own product? My team at aishifu offers this as a service: aishifu.shop/en/ad-remake/.