What an AI product video really costs when you count the retries

The full ledger: 1,890 credits, not one 330-credit render

The product-only run used an existing AI-generated candle photo. It charged 330 credits for the video and incurred no paid reference-image generation in that run. The photo still had a history: saying there was no new reference charge does not mean producing every possible starting asset is free. The public record begins with the existing image supplied to this test.

The presenter workflow was a separate sequence. Three generated reference images used 900 credits. Its first video used 330 credits and was rejected for framing. A corrected video used another 330 credits. The presenter sequence therefore cost 1,560 credits. Add the separate product-only clip and the recorded total becomes 1,890. These numbers describe our dated test, not a universal rate card.

The public product example and presenter example show the files and limitations. Keep those workflows separate when quoting the figures: 1,560 is the presenter sequence; 1,890 includes the product-only test too.

• Product-only video: 330 credits; existing candle photo; five-second raw output.
• Presenter reference images: 900 credits across three generated images.
• First presenter video: 330 credits; rejected after review.
• Presenter correction: 330 credits; framing improved, residual mouth movement remained.
• Presenter subtotal: 900 + 330 + 330 = 1,560 credits. All recorded runs: 330 + 1,560 = 1,890 credits.

What the product-only clip delivered

The source showed an amber candle jar, a blank kraft label and a bedroom setting. The raw video was 720 by 1296 pixels and about 5.04 seconds long. Five sampled frames kept the candle and surrounding props recognizable. That is a limited visual observation, not a guarantee that every edge, dimension or label position matches a real item.

The requested brief called for one continuous shot and no sound. The expanded request introduced two timeline beats and a cut, and the output visibly changed framing around the third second. The raw MP4 also contained an audio track. Its automatic cut added music and Chinese text. We did not accept that automatic cut as a finished advertisement.

This distinction matters when comparing prices. A generated file can be a useful production input while still failing a delivery requirement. Record the output you actually reviewed, including whether it was the raw model file or an assembled export. A 1080 by 1920 automatic export does not prove that the underlying generation contains native 1080p detail.

Why the presenter sequence needed a correction

The first presenter attempt used generated references and produced a split layout around the third second. It also showed a speaking mouth. The automatic assembly added captions and music. Those details conflicted with the intended silent, single-scene demonstration, so the 330-credit attempt stayed in the ledger as rejected work. Deleting it from a portfolio would not delete its production cost.

For the correction, we supplied one generated start frame containing the candle and fictional presenter. We removed the style reference and storyboard grid, and switched Sound and Prompt enhance off. The corrected raw file was five seconds long, 720 by 1296 pixels and had no audio stream. Five sampled frames showed one full-frame scene with the same presenter and candle.

The correction still showed unwanted mouth and hand movement. Silence is a property of the file, not proof that the face followed every instruction. We describe this as an internal demonstration with a partial correction. It supplies no evidence about voice quality, lip synchronization, real customer experience or advertising approval. A budget should preserve that qualified outcome instead of calling the retry a complete success.

Copy a ledger that includes review work

Start one record per attempt before you press Generate. Give the brief an identifier, save the source photo and record which product variant it shows. A later reviewer should be able to distinguish a corrected prompt from a new creative idea. Otherwise a cheaper variation can accidentally be compared with a much more demanding delivery brief.

The fields below are enough for a small team to reconstruct a decision. Track review time separately from credits: our public test records do not contain measured labor hours, so this article assigns no labor price. If a designer spends time repairing a label or removing sound, add their actual time and cost in your own ledger rather than treating that repair as part of the model output.

Also separate a reusable reference from a one-use expense. You may later reuse a presenter image across several clips, but our first sequence paid the entire 900-credit image charge before either video was reviewed. Report both the outlay for the batch and any allocation you use per deliverable. An allocation should not conceal the original expense.

• Brief ID and date; exact product variant; source-file location; reference images already owned or newly generated.
• Attempt number; model and settings; requested duration and ratio; quoted credits; actual charged credits.
• Output-file location; actual dimensions and duration; audio present or absent; raw or assembled export.
• Review outcome: usable as specified, needs correction, or rejected; precise reason; reviewer and time spent.
• Correction link; additional charges; final use; total batch spend and number of deliverables accepted for that use.

Define acceptance before choosing a cheaper render

A useful acceptance rule describes something you can see or inspect. For a candle demonstration that could mean one jar throughout, the correct label, no hands near the flame, one continuous scene and no audio stream. A request for a beautiful video is too loose to explain why a take should be accepted or rejected. Write the conditions beside the brief before seeing the result.

Watch the complete clip, then inspect the beginning, middle, ending and any visible transition. Compare the product with the supplied photograph. Read any lettering at a size close to its eventual placement. Check the downloaded file as well as the preview, since an automatic assembly may add captions, music or a different export size. Keep a note when a condition was only sampled rather than verified throughout.

The checks should match the destination. A product listing needs an accurate item, while an ad also needs a clear offer and a readable ending. The TikTok safe-zone checker helps inspect placement, but it does not establish product fidelity or approve advertising claims. Passing one check does not remove the need to inspect the rest of the creative.

Compare total production cost with the same denominator

Cost per render answers how much an attempt used. Cost per accepted deliverable answers how much production spend led to something your team can use. Calculate the latter as total relevant spend divided by deliverables accepted for the stated purpose. Include rejected attempts and reference creation. Keep human review, editing, subscriptions and media buying in separate, named rows so the comparison is understandable.

Our candle record does not establish a cost per approved ad: the public clips retained unmet constraints, and no ad approval was measured. Dividing 1,890 by three and calling the result a successful video price would hide that fact. Dividing only 330 by a corrected clip would also omit the references and rejected take that preceded it. The honest result here is the recorded workflow cost and its documented outcome.

When comparing tools, use the same source photo, brief, length, review conditions and number of variations. Record dates because models, settings and commercial terms can change. Credits are a product accounting unit, not automatically comparable dollars across services. For your own Blurbly budget, check the current plan and the quoted generation cost rather than converting this historical ledger with an assumed credit price.

Use the next attempt to answer one specific question

When a take fails, write the failure in a sentence before retrying. In our presenter sequence, the framing problem led to a single start-frame input rather than another storyboard grid. That change gave us a clear question to review: did the output remain in one full-frame scene? It did in the sampled frames, while mouth movement remained unresolved.

For your own product, decide whether the remaining problem justifies another render, a manual edit or a different creative concept. Set a batch budget before expanding into several hooks. Keep the original files and review notes, including failures, so the next person can understand what the spend taught you. A small, documented experiment can inform a larger production decision without pretending to prove advertising performance.