Editing AI-generated video means assembling short clips into one timed cut. Every model in a production catalog returns a few seconds per generation, so a 15 or 30 second ad is four to six clips joined on a timeline, trimmed, captioned and scored. The edit decides pacing, sound and aspect ratio. The prompt never does.
You wrote a good prompt. The model returned eight seconds of a bag turning on a plinth, and it looks right. Then you open the placement brief. The ad is 30 seconds, it has to read with the sound off, and it ships in four shapes.
That gap is the step nobody writes about: the assembly. This guide covers what the clip length limits are across the models, what a finished ad needs on top of raw generations, and the order that stops you regenerating work the timeline could have fixed.
Key Takeaways
- No model gives you a finished ad. Most generations run 4 to 10 seconds, and Seedance 2.0 and Kling 3.0 reach 15.
- A 30-second spot is four to six generations. Budget the clip count before you budget the credits.
- Captions are a high-return edit. Meta lists them as recommended on its own Feed spec, and they cost nothing to add on the timeline.
- Fix it in the edit before you regenerate. Trimming a soft ending is free. Re-rolling the clip is not.
- One master cut, four shapes. Take 16:9, 9:16, 1:1 and 4:5 from the same cut instead of four separate prompt sessions.
- The widely quoted “85% watch without sound” figure came from publishers in 2016, and Meta did not publish it. The better-sourced survey number is lower and several years old. The direction still holds.
Why every AI clip arrives too short
Generation length is capped per model, and the caps are tighter than most briefs assume. These are the durations and ratios each vendor documents, as of September 2026.
| Model | Clip durations offered | Aspect ratios |
|---|---|---|
| Veo 3.1 | 4s, 6s, 8s | 16:9, 9:16 |
| Sora 2 Pro | Up to 20s in OpenAI’s guide. Removed from OpenAI’s API on 24 September 2026 | 16:9, 9:16 |
| Seedance 2.0 | 4s to 15s, any whole second | 16:9, 4:3, 1:1, 3:4, 9:16, 21:9 |
| Kling 2.6 Pro | 5s or 10s | 16:9, 9:16, 1:1 (text to video) |
| Runway Gen-4.5 | 2s to 10s, any whole second | 16:9, 9:16, 1:1, 4:3, 3:4, 21:9 (image to video) |
OpenAI is removing Sora 2 and Sora 2 Pro from its API on 24 September 2026 (OpenAI API deprecations, accessed September 2026). Until then, OpenAI’s video guide lists Sora 2 Pro clips of up to 20 seconds, while its API reference still lists 4, 8 and 12. For a take longer than 8 seconds after that date, Seedance 2.0 runs 4 to 15 seconds (BytePlus docs, September 2026) and Kling 3.0 runs 3 to 15 seconds (Kling capability map, September 2026). Kling 2.6 makes 5 or 10 second clips. Runway Gen-4.5 takes any whole number of seconds from 2 to 10 (Runway API docs, September 2026).
The caps come from the models themselves, not from the plan. Google’s own documentation sets Veo 3.1 at 4, 6 or 8 seconds per generation, and requires the 8-second setting for 1080p and 4K output (ai.google.dev, September 2026).
Veo does support extension, and the limits there are worth reading before you plan around it. In the Gemini API, Google documents extending a Veo clip by 7 seconds at a time, up to 20 times, for up to 148 seconds. Extension in the Gemini API runs at 720p only, and Google Cloud documents a different, shorter limit (ai.google.dev, September 2026). So the long-form route exists, and it costs you resolution.
For an ad, extension is usually the wrong tool anyway. A 30-second commercial is six shots.
What a 30-second ad needs
Take the standard shape of a direct-response spot: hook, product, benefit, proof, offer, end card. That is six beats. At 4 to 8 seconds a generation, it is six generations, and the six were never going to match on their own.
Here is what the assembly adds that no single generation gives you:
- Cuts on the beat. Raw clips run to their full length. Almost every one is stronger 1 to 2 seconds shorter.
- Order. The shot you generated third is often the hook.
- Captions. Text that carries the claim when the sound is off.
- Sound. A bed, a voiceover, or both. Native model audio rarely matches across six separate clips.
- The end card. The logo, the offer and the call to action are graphics you place on the timeline.
- The export set. Four ratios from one cut.
The timeline does all six jobs.
The five jobs the timeline does
The DesignerBox video editor is a real timeline, with several tracks, transitions, animated text and audio. That matters when you are matching a brief instead of chasing a look. It also handles picture in picture and LUT colour looks.
The five jobs, in the order they pay off:
Trim. The cheapest quality gain available. Generated clips tend to soften in the last beat as the motion resolves. Cut before it does.
Sequence. Reordering costs nothing and changes the hook. Try the product reveal first before you accept the slow build.
Caption. Put the claim on screen as animated text, timed to the shot it describes.
Score. Put the music bed and the voiceover on audio tracks under the picture. DesignerBox also has text to speech with 20 voices for a voiceover. Check the licence terms of any music you add.
Grade. Apply one colour look across every clip. That one look is what makes six separate generations read as one shoot.
Captions, and the statistic everyone gets wrong
You have read that 85% of Facebook video is watched without sound. It is worth knowing where that number comes from, because it is not from Meta.
It traces to 2016 reporting in Digiday, which quoted individual publishers describing their own traffic. LittleThings said roughly 85% of its Facebook viewership happened without sound, and Mic reported a similar share of its 30-second views. Those are publisher figures for publisher content from a decade ago, not a platform-wide measurement of ads.
The better-sourced number is a Verizon Media and Publicis Media survey of 5,616 US adults, reported in 2019: 69% said they watch video with the sound off in public places, and 80% said they were more likely to finish a video when captions were available.
Both numbers are old. Neither is a controlled test of your ad. What holds up is the platform’s own guidance: Meta’s ads guide lists both sound and subtitles as optional but recommended on its Feed video spec (Meta Ads Guide, September 2026).
Caption anyway. It is a short pass on the timeline, and it lets a silent viewer follow the claim.
Export once, ship four shapes
Meta’s published Feed video spec is 4:5 at 1440 x 1800 pixels, MP4, MOV or GIF, from 1 second to 241 minutes, up to 4 GB (Meta Ads Guide, September 2026). Stories and Reels are full-screen vertical, and a YouTube in-stream spot is usually 16:9. A cut framed for one shape needs reframing for the others.
Generating each shape separately is the expensive mistake. Two 9:16 generations of the same brief will not match, because you are rolling the dice twice. Cut the master once, then export the ratios, and the shots stay identical across placements.
For the same step on ad creative, the ads resizer app makes one ad in every size the ad platforms need.
Where the edit belongs in the production order
Put the edit before the second generation round, not after it.
The instinct is to keep prompting until the clips are perfect, then assemble. That order pays for problems the timeline solves for free. Trim a clip that ends weakly instead of re-rolling it. Reorder a cut that feels slow before you make new footage. Fix a shot that reads too cool with the grade you apply to all six, and keep the prompt as it is.
The working order:
- Generate one clip per beat, cheapest usable setting.
- Assemble a rough cut and watch it once, silent.
- Trim, reorder, grade.
- List only the shots that still fail. Regenerate those.
- Caption, score, export.
Step 4 is usually one or two clips, not six. That is the whole saving, and it lands on your credit balance directly, since video is priced by length and every re-roll adds to the bill. In DesignerBox, an 8-second clip costs 40 to 560 credits, depending on the model, and the cost is shown before the run.
What this changes about how you prompt
Once you plan for an edit, you prompt differently.
Ask for one action per clip, not three. A clip that tries to pan, reveal and cut inside six seconds gives the timeline nothing to work with. Six clean single-action clips cut into a better ad than three busy ones.
Leave handles. Generate slightly longer than the beat needs so you have material to trim into.
Stop asking the model for text. On-screen type belongs on the timeline, where you can edit a price without paying for a new generation, and where it comes out sharp.
Match the source. Shots derived from the same product photo cut together far more easily than shots built from separate text prompts, which is the same reason a product photo makes a better source for a video ad than a written brief.
If you are planning this for paid social at volume, the arithmetic of how many clips a campaign needs is worked through in the guide to AI video ad packs, and the platform-specific version for short-form sits in how to create TikTok video ads with AI.
The finished cut in DesignerBox
Make one clip per beat, assemble them in the video editor, caption the cut with animated text and download the result. If you would rather start from a product photo than from a blank timeline, a product-to-video ad template is a workflow somebody already built. Open it, add your photo and run it.
The model is one step in the workflow. You can pick a different model for each shot, so the hook gets a higher-cost model and the filler beats a lower-cost one. The model list shows the 13 video models.
Keep the steps as a workflow for the next product. You set the brand once, and the workflow reads it on every run, so the next product gets the same model, light and framing. Batch is coming, which will run one workflow over a whole range. Start from the templates and run your first product. The cost is shown before the run.
FAQ
Can AI generate a full 30-second ad in one clip?
Not as a single generation from the models in the DesignerBox catalog. Sora 2 Pro makes up to 20 seconds until 24 September 2026, the day OpenAI removes it from its API. The longest single generation among the other catalog models is 15 seconds, on Seedance 2.0 and Kling 3.0 (docs.byteplus.com, kling.ai, September 2026). Outside the catalog, ByteDance’s Seedance 2.5 runs up to 30 seconds (seed.bytedance.com, September 2026). A 30-second ad from the catalog models is assembled from several generations on a timeline.
How many clips does a 30-second ad need?
Four to six for a standard direct-response structure: hook, product, benefit, proof, offer, end card. Generate one clip per beat, then trim each by a second or two in the edit.
Do AI video clips come with usable audio?
Some models generate native audio, but ad audio usually needs a music bed, a voiceover or both, mixed to a consistent level across shots. That is a timeline job. The DesignerBox video editor has audio tracks, and DesignerBox has text to speech with 20 voices for the voiceover.
Should I add captions to AI-generated video ads?
Yes. Meta lists subtitles as recommended on its Feed video spec (Meta Ads Guide, September 2026), and captions let a silent viewer follow the claim. In the DesignerBox video editor, you add them as animated text on the timeline.
Can I export one edit to every placement size?
Yes, in most editors. Cut one master, then reframe it for 16:9, 9:16, 1:1 and 4:5, and the shots stay identical across placements. Frame each shot with space around the subject so every crop still holds it. For static ad creative, the DesignerBox ads resizer makes one ad in every size the ad platforms need.
Is it cheaper to fix a clip in the edit or regenerate it?
Fix it in the edit where you can. Trimming, reordering and grading cost nothing. Regeneration is priced by clip length, so re-rolling six clips to solve a pacing problem spends credits on something the timeline already handles.
What resolution should I export ad video at?
Check each placement’s own spec. Meta’s Feed spec recommends 1440 x 1800 for 4:5 (Meta Ads Guide, September 2026), so export at least 1080p and match the recommended size where you can. Veo’s extension in the Gemini API runs at 720p only, so long-form built that way will not reach 1080p (ai.google.dev, September 2026).
Sources
- Veo 3.1 clip durations, resolution requirements and Gemini API extension limits: ai.google.dev, accessed September 2026
- Sora 2 and Sora 2 Pro removal from the API on 24 September 2026: OpenAI API deprecations, accessed September 2026
- Sora 2 Pro durations and sizes: OpenAI video generation guide, accessed September 2026
- Seedance 2.0 durations and aspect ratios: docs.byteplus.com, accessed September 2026
- Seedance 2.5 durations: seed.bytedance.com, accessed September 2026
- Kling 2.6 and Kling 3.0 durations and ratios: kling.ai capability map, accessed September 2026
- Runway Gen-4.5 durations: docs.dev.runwayml.com, and aspect ratios: help.runwayml.com, accessed September 2026
- DesignerBox video editor (designerbox.ai/video-editor), September 2026: several tracks, picture in picture, transitions, animated text, audio and LUT; text to speech with 20 voices; the ads resizer
- Meta Feed video ad specification and caption guidance: facebook.com, accessed September 2026
- Sound-off viewing figures: Digiday, reporting publisher-supplied data, May 2016; Verizon Media and Publicis Media survey of 5,616 US adults, reported by Forbes, July 2019; both accessed September 2026
Model clip durations checked against vendor documentation, and editor capabilities against the DesignerBox video editor page, as of September 2026. Sound-off viewing figures are third-party survey data from 2016 and 2019 and are cited with those dates.
