Fitting CogVideoX output into YouTube ad formats
The CogVideoX-5B card states a length of six seconds at 49 frames, a resolution of 720 x 480 with no support for other resolutions, and eight frames per second. Every figure is a single value, which makes the arithmetic against an ad slot exact and the room to negotiate nil. As of 2026-09-22.
| Format | Generations needed | Note |
|---|---|---|
| Bumper, 5-6 s | One generation | Six seconds is the published length |
| Non-skippable, up to 60 s | Ten generations | The length is fixed, so the count is exact |
| Skippable, 30 s of paid view | Five generations | No shorter or longer run is offered |
| Masthead, 16:9 up to 30 s | Not stated | 720 x 480, with no other resolution supported |
Inclusion rule. Formats from Google's own help page, matched only against limits the tool vendor publishes. Where either side does not state a value, the row says so rather than estimating. Order. Shortest format first, as on the placements page.
1How much each tool puts in writing
| Tool | Limits published |
|---|---|
| Amazon Nova Reel | 4 published |
| CogVideoX | 3 published |
| Google Veo | 4 published |
| Grok Imagine | 4 published |
| Hedra | 3 published |
| HeyGen | 4 published |
| LTX-Video | 3 published |
| MiniMax | 3 published |
| Pika | 2 published |
| PixVerse | 3 published |
| Runway | 4 published |
| SceneMixer | 3 published |
| Wan 3.0 | 4 published |
Inclusion rule. Tools with a sheet on this site, counted by how many separate output limits their own pages state. A limit mentioned only in a launch post is not counted. Order. Alphabetical by tool name.
| Tool | Bumper, 5-6 s |
|---|---|
| Amazon Nova Reel | One shot |
| CogVideoX | One generation |
| Google Veo | One generation |
| Grok Imagine | One generation |
| Hedra | One generation |
| HeyGen | One scene |
| LTX-Video | One generation |
| MiniMax | One generation |
| Pika | One generation |
| PixVerse | One generation |
| Runway | One generation |
| SceneMixer | One generation |
| Wan 3.0 | One generation |
Inclusion rule. Tools publishing output limits on their own product pages. A tool silent on this particular limit keeps its row, marked hollow. Order. Alphabetical by tool name.
| Tool | Non-skippable, up to 60 s |
|---|---|
| Amazon Nova Reel | One request of ten shots |
| CogVideoX | Ten generations |
| Google Veo | Eight generations |
| Grok Imagine | Four generations |
| Hedra | One generation |
| HeyGen | One scene |
| LTX-Video | Seven generations |
| MiniMax | Four generations |
| Pika | Six or more generations |
| PixVerse | Four generations |
| Runway | Four generations |
| SceneMixer | Two generations on Wan 3.0 |
| Wan 3.0 | Two generations |
Inclusion rule. Tools publishing output limits on their own product pages. A tool silent on this particular limit keeps its row, marked hollow. Order. Alphabetical by tool name.
| Tool | Skippable, 30 s of paid view |
|---|---|
| Amazon Nova Reel | One request of five shots |
| CogVideoX | Five generations |
| Google Veo | Four generations |
| Grok Imagine | Two generations |
| Hedra | One generation |
| HeyGen | One scene |
| LTX-Video | Four generations |
| MiniMax | Two generations |
| Pika | Two to three generations |
| PixVerse | Two generations |
| Runway | Two generations |
| SceneMixer | One generation on Wan 3.0 |
| Wan 3.0 | One generation |
Inclusion rule. Tools publishing output limits on their own product pages. A tool silent on this particular limit keeps its row, marked hollow. Order. Alphabetical by tool name.
2A specification with no ranges in it
Most entries on these sheets publish a span or a menu, and the interesting question is which rung to buy. This card publishes three numbers and closes the question: six seconds, 720 by 480, eight frames per second. The phrase that matters is the parenthetical after the resolution, which says other resolutions are not supported rather than not recommended.
For a media plan that is oddly useful. A six second bumper is one run with no decision in it, and a sixty second non-skippable is ten runs and nine joins, exactly, before anyone opens an editor. Nothing about the count depends on a setting a producer might get wrong.
3Eight frames a second is the figure that changes the brief
Eight frames per second is roughly a third of what a viewer expects from moving picture, and the card states it as the frame rate rather than as a sampling step. A piece delivered at that rate reads as a sequence of held images with movement between them, which suits some kinds of advertising and ruins any shot whose point is speed.
Interpolating up to 24 is the obvious response, and it is a step this register cannot record because the card does not offer it. What the card does offer is an honest statement of what leaves the model, which is more than a page listing a resolution tier and saying nothing about how many pictures a second arrive in it.
4Sources
Output limits read from the model card at huggingface.co, checked 2026-09-22. Format requirements are from Google Ads Help, and the matching rule is on method. Next along the register: Google Veo, Grok Imagine. The whole set of 15 is on published limits, and the sheets for each are indexed at tools.
- Inference resolution720 x 480, no support for other resolutionsfrom the model card table
- Video length6 seconds, at 49 framesfrom the model card table
- Frame rate8 frames per secondfrom the model card table